FROM vllm/vllm-openai:v0.8.1

ARG MODEL_DIR=”models/Qwen2.5-VL-7B-Instruct”

RUN pip install “vllm
==0.8.1″ accelerate qwen-vl-utils

RUN mkdir -p /model-directory

COPY ${MODEL_DIR} /model-directory/Qwen2.5-VL-7B-Instruct

ENV NCCL_SHM_DISABLE=1

ENTRYPOINT [“python3”, “-m”, “vllm.entrypoints.openai.api_server”, \
“–model”, “/model-directory/Qwen2.5-VL-7B-Instruct/”, \
“–dtype”, “bfloat16”, \
“–enable-prefix-caching”, \
“–served-model-name”, “qwen-vl”, \
“–gpu_memory_utilization”, “0.75”, \
“–max-model-len”, “32768”, \
“–limit-mm-per-prompt”, “image=2,video=1”]
Post10

Post7

By cspoweruser on

June 14, 2025

- Uncategorized

Gemma 3n model card

Model Page: Gemma 3n

Resources and Technical Documentation:

Responsible Generative AI Toolkit
Gemma on Kaggle
Google AI Edge documentation to run on mobile
Try on Android by downloading our Google AI Edge Gallery sample app

Terms of Use: Terms
Authors: Google DeepMind

Model Information

Summary description and brief definition of inputs and outputs.

Description

Gemma is a family of lightweight, state-of-the-art open models from Google, built from the same research and technology used to create the Gemini models. Gemma models are well-suited for a variety of content understanding tasks, including question answering, summarization, and reasoning. Their relatively small size makes it possible to deploy them in environments with limited resources such as laptops, desktops or your own cloud infrastructure, democratizing access to state of the art AI models and helping foster innovation for everyone.

Gemma 3n models are designed for efficient execution on low-resource devices. They are capable of multimodal input, handling text, image, video, and audio input, and generating text outputs, with open weights for instruction-tuned variants. These models were trained with data in over 140 spoken languages.

Gemma 3n models use selective parameter activation technology to reduce resource requirements. This technique allows the models to operate at an effective size of 2B and 4B parameters, which is lower than the total number of parameters they contain. For more information on Gemma 3n’s efficient parameter management technology, see the Gemma 3n page.

Inputs and outputs

Input:
- Text string, such as a question, a prompt, or a document to be summarized
- Images, normalized to 256×256, 512×512, or 768×768 resolution and encoded to 256 tokens each
- Audio data encoded to 6.25 tokens per second from a single channel
- Total input context of 32K tokens
Output:
- Generated text in response to the input, such as an answer to a question, analysis of image content, or a summary of a document
- Total output length up to 32K tokens, subtracting the request input tokens

Citation

@article{gemma_3n_2025,
    title={Gemma 3n},
    url={https://ai.google.dev/gemma/docs/gemma-3n},
    publisher={Google DeepMind},
    author={Gemma Team},
    year={2025}
}

Model Data

Data used for model training and how the data was processed.

Training Dataset

These models were trained on a dataset that includes a wide variety of sources totalling approximately 11 trillion tokens. The knowledge cutoff date for the training data was June 2024. Here are the key components:

Web Documents: A diverse collection of web text ensures the model is exposed to a broad range of linguistic styles, topics, and vocabulary. The training dataset includes content in over 140 languages.
Code: Exposing the model to code helps it to learn the syntax and patterns of programming languages, which improves its ability to generate code and understand code-related questions.
Mathematics: Training on mathematical text helps the model learn logical reasoning, symbolic representation, and to address mathematical queries.
Images: A wide range of images enables the model to perform image analysis and visual data extraction tasks.
Audio: A diverse set of sound samples enables the model to recognize speech, transcribe text from recordings, and identify information in audio data.

The combination of these diverse data sources is crucial for training a powerful multimodal model that can handle a wide variety of different tasks and data formats.

Data Preprocessing

Here are the key data cleaning and filtering methods applied to the training data:

CSAM Filtering: Rigorous CSAM (Child Sexual Abuse Material) filtering was applied at multiple stages in the data preparation process to ensure the exclusion of harmful and illegal content.
Sensitive Data Filtering: As part of making Gemma pre-trained models safe and reliable, automated techniques were used to filter out certain personal information and other sensitive data from training sets.
Additional methods: Filtering based on content quality and safety in line with our policies.

Implementation Information

Details about the model internals.

Hardware

Gemma was trained using Tensor Processing Unit (TPU) hardware (TPUv4p, TPUv5p and TPUv5e). Training generative models requires significant computational power. TPUs, designed specifically for matrix operations common in machine learning, offer several advantages in this domain:

Performance: TPUs are specifically designed to handle the massive computations involved in training generative models. They can speed up training considerably compared to CPUs.
Memory: TPUs often come with large amounts of high-bandwidth memory, allowing for the handling of large models and batch sizes during training. This can lead to better model quality.
Scalability: TPU Pods (large clusters of TPUs) provide a scalable solution for handling the growing complexity of large foundation models. You can distribute training across multiple TPU devices for faster and more efficient processing.
Cost-effectiveness: In many scenarios, TPUs can provide a more cost-effective solution for training large models compared to CPU-based infrastructure, especially when considering the time and resources saved due to faster training.

These advantages are aligned with Google’s commitments to operate sustainably.

Software

Training was done using JAX and ML Pathways. JAX allows researchers to take advantage of the latest generation of hardware, including TPUs, for faster and more efficient training of large models. ML Pathways is Google’s latest effort to build artificially intelligent systems capable of generalizing across multiple tasks. This is specially suitable for foundation models, including large language models like these ones.

Together, JAX and ML Pathways are used as described in the paper about the Gemini family of models: “the ‘single controller’ programming model of Jax and Pathways allows a single Python process to orchestrate the entire training run, dramatically simplifying the development workflow.”

Evaluation

Model evaluation metrics and results.

Benchmark Results

These models were evaluated at full precision (float32) against a large collection of different datasets and metrics to cover different aspects of content generation. Evaluation results marked with IT are for instruction-tuned models. Evaluation results marked with PT are for pre-trained models.

Reasoning and factuality

Benchmark	Metric	n-shot	E2B PT	E4B PT
HellaSwag	Accuracy	10-shot	72.2	78.6
BoolQ	Accuracy	0-shot	76.4	81.6
PIQA	Accuracy	0-shot	78.9	81.0
SocialIQA	Accuracy	0-shot	48.8	50.0
TriviaQA	Accuracy	5-shot	60.8	70.2
Natural Questions	Accuracy	5-shot	15.5	20.9
ARC-c	Accuracy	25-shot	51.7	61.6
ARC-e	Accuracy	0-shot	75.8	81.6
WinoGrande	Accuracy	5-shot	66.8	71.7
BIG-Bench Hard	Accuracy	few-shot	44.3	52.9
DROP	Token F1 score	1-shot	53.9	60.8

Multilingual

Benchmark	Metric	n-shot	E2B IT	E4B IT
MGSM	Accuracy	0-shot	53.1	60.7
WMT24++ (ChrF)	Character-level F-score	0-shot	42.7	50.1
Include	Accuracy	0-shot	38.6	57.2
MMLU (ProX)	Accuracy	0-shot	8.1	19.9
OpenAI MMLU	Accuracy	0-shot	22.3	35.6
Global-MMLU	Accuracy	0-shot	55.1	60.3
ECLeKTic	ECLeKTic score	0-shot	2.5	1.9

STEM and code

Benchmark	Metric	n-shot	E2B IT	E4B IT
GPQA Diamond	RelaxedAccuracy/accuracy	0-shot	24.8	23.7
LiveCodeBench v5	pass@1	0-shot	18.6	25.7
Codegolf v2.2	pass@1	0-shot	11.0	16.8
AIME 2025	Accuracy	0-shot	6.7	11.6

Additional benchmarks

Benchmark	Metric	n-shot	E2B IT	E4B IT
MMLU	Accuracy	0-shot	60.1	64.9
MBPP	pass@1	3-shot	56.6	63.6
HumanEval	pass@1	0-shot	66.5	75.0
LiveCodeBench	pass@1	0-shot	13.2	13.2
HiddenMath	Accuracy	0-shot	27.7	37.7
Global-MMLU-Lite	Accuracy	0-shot	59.0	64.5
MMLU (Pro)	Accuracy	0-shot	40.5	50.6

Android Performance Benchmarks with Google AI Edge

Note that all benchmark stats are from a Samsung S25 Ultra with 4096 KV cache size, 1024 tokens prefill, 256 tokens decode.

These numbers will continue to improve while Gemma 3n is in preview.

Weight Quantization	Backend	Prefill (tokens/sec)	Decode (tokens/sec)	Time to first token (sec)	Model size (MB)	Peak RSS Memory (MB)	GPU Memory (MB)
dynamic_int4	CPU	118	12.8	9.2	4201	3924	193
dynamic_int4	GPU	446	16.1	15.1	4201	5504	3048

Model size: measured by the size of the .tflite flatbuffer (serialization format for LiteRT models)
The inference on CPU is accelerated via the LiteRT XNNPACK delegate with 4 threads
Benchmark on CPU is done assuming XNNPACK cache is enabled
Benchmark on GPU is done assuming model is cached
Vision encoder is always run on GPU with 512×512 resolution
Cpufreq governor is set to performance during benchmark. Observed performance may vary depending on your phone’s hardware and current activity level.
dynamic_int4: quantized model with int4 weights and float activations.

0

Post6

By cspoweruser on

June 14, 2025

- Uncategorized

Gemma 3n model card

Model Page: Gemma 3n

Resources and Technical Documentation:

Responsible Generative AI Toolkit
Gemma on Kaggle
Google AI Edge documentation to run on mobile
Try on Android by downloading our Google AI Edge Gallery sample app

Terms of Use: Terms
Authors: Google DeepMind

Model Information

Summary description and brief definition of inputs and outputs.

Description

Gemma is a family of lightweight, state-of-the-art open models from Google, built from the same research and technology used to create the Gemini models. Gemma models are well-suited for a variety of content understanding tasks, including question answering, summarization, and reasoning. Their relatively small size makes it possible to deploy them in environments with limited resources such as laptops, desktops or your own cloud infrastructure, democratizing access to state of the art AI models and helping foster innovation for everyone.

Gemma 3n models are designed for efficient execution on low-resource devices. They are capable of multimodal input, handling text, image, video, and audio input, and generating text outputs, with open weights for instruction-tuned variants. These models were trained with data in over 140 spoken languages.

Gemma 3n models use selective parameter activation technology to reduce resource requirements. This technique allows the models to operate at an effective size of 2B and 4B parameters, which is lower than the total number of parameters they contain. For more information on Gemma 3n’s efficient parameter management technology, see the Gemma 3n page.

Inputs and outputs

Input:
- Text string, such as a question, a prompt, or a document to be summarized
- Images, normalized to 256×256, 512×512, or 768×768 resolution and encoded to 256 tokens each
- Audio data encoded to 6.25 tokens per second from a single channel
- Total input context of 32K tokens
Output:
- Generated text in response to the input, such as an answer to a question, analysis of image content, or a summary of a document
- Total output length up to 32K tokens, subtracting the request input tokens

Citation

@article{gemma_3n_2025,
    title={Gemma 3n},
    url={https://ai.google.dev/gemma/docs/gemma-3n},
    publisher={Google DeepMind},
    author={Gemma Team},
    year={2025}
}

Model Data

Data used for model training and how the data was processed.

Training Dataset

These models were trained on a dataset that includes a wide variety of sources totalling approximately 11 trillion tokens. The knowledge cutoff date for the training data was June 2024. Here are the key components:

Web Documents: A diverse collection of web text ensures the model is exposed to a broad range of linguistic styles, topics, and vocabulary. The training dataset includes content in over 140 languages.
Code: Exposing the model to code helps it to learn the syntax and patterns of programming languages, which improves its ability to generate code and understand code-related questions.
Mathematics: Training on mathematical text helps the model learn logical reasoning, symbolic representation, and to address mathematical queries.
Images: A wide range of images enables the model to perform image analysis and visual data extraction tasks.
Audio: A diverse set of sound samples enables the model to recognize speech, transcribe text from recordings, and identify information in audio data.

The combination of these diverse data sources is crucial for training a powerful multimodal model that can handle a wide variety of different tasks and data formats.

Data Preprocessing

Here are the key data cleaning and filtering methods applied to the training data:

CSAM Filtering: Rigorous CSAM (Child Sexual Abuse Material) filtering was applied at multiple stages in the data preparation process to ensure the exclusion of harmful and illegal content.
Sensitive Data Filtering: As part of making Gemma pre-trained models safe and reliable, automated techniques were used to filter out certain personal information and other sensitive data from training sets.
Additional methods: Filtering based on content quality and safety in line with our policies.

Implementation Information

Details about the model internals.

Hardware

Gemma was trained using Tensor Processing Unit (TPU) hardware (TPUv4p, TPUv5p and TPUv5e). Training generative models requires significant computational power. TPUs, designed specifically for matrix operations common in machine learning, offer several advantages in this domain:

Performance: TPUs are specifically designed to handle the massive computations involved in training generative models. They can speed up training considerably compared to CPUs.
Memory: TPUs often come with large amounts of high-bandwidth memory, allowing for the handling of large models and batch sizes during training. This can lead to better model quality.
Scalability: TPU Pods (large clusters of TPUs) provide a scalable solution for handling the growing complexity of large foundation models. You can distribute training across multiple TPU devices for faster and more efficient processing.
Cost-effectiveness: In many scenarios, TPUs can provide a more cost-effective solution for training large models compared to CPU-based infrastructure, especially when considering the time and resources saved due to faster training.

These advantages are aligned with Google’s commitments to operate sustainably.

Software

Training was done using JAX and ML Pathways. JAX allows researchers to take advantage of the latest generation of hardware, including TPUs, for faster and more efficient training of large models. ML Pathways is Google’s latest effort to build artificially intelligent systems capable of generalizing across multiple tasks. This is specially suitable for foundation models, including large language models like these ones.

Together, JAX and ML Pathways are used as described in the paper about the Gemini family of models: “the ‘single controller’ programming model of Jax and Pathways allows a single Python process to orchestrate the entire training run, dramatically simplifying the development workflow.”

Evaluation

Model evaluation metrics and results.

Benchmark Results

These models were evaluated at full precision (float32) against a large collection of different datasets and metrics to cover different aspects of content generation. Evaluation results marked with IT are for instruction-tuned models. Evaluation results marked with PT are for pre-trained models.

Reasoning and factuality

Benchmark	Metric	n-shot	E2B PT	E4B PT
HellaSwag	Accuracy	10-shot	72.2	78.6
BoolQ	Accuracy	0-shot	76.4	81.6
PIQA	Accuracy	0-shot	78.9	81.0
SocialIQA	Accuracy	0-shot	48.8	50.0
TriviaQA	Accuracy	5-shot	60.8	70.2
Natural Questions	Accuracy	5-shot	15.5	20.9
ARC-c	Accuracy	25-shot	51.7	61.6
ARC-e	Accuracy	0-shot	75.8	81.6
WinoGrande	Accuracy	5-shot	66.8	71.7
BIG-Bench Hard	Accuracy	few-shot	44.3	52.9
DROP	Token F1 score	1-shot	53.9	60.8

Multilingual

Benchmark	Metric	n-shot	E2B IT	E4B IT
MGSM	Accuracy	0-shot	53.1	60.7
WMT24++ (ChrF)	Character-level F-score	0-shot	42.7	50.1
Include	Accuracy	0-shot	38.6	57.2
MMLU (ProX)	Accuracy	0-shot	8.1	19.9
OpenAI MMLU	Accuracy	0-shot	22.3	35.6
Global-MMLU	Accuracy	0-shot	55.1	60.3
ECLeKTic	ECLeKTic score	0-shot	2.5	1.9

STEM and code

Benchmark	Metric	n-shot	E2B IT	E4B IT
GPQA Diamond	RelaxedAccuracy/accuracy	0-shot	24.8	23.7
LiveCodeBench v5	pass@1	0-shot	18.6	25.7
Codegolf v2.2	pass@1	0-shot	11.0	16.8
AIME 2025	Accuracy	0-shot	6.7	11.6

Additional benchmarks

Benchmark	Metric	n-shot	E2B IT	E4B IT
MMLU	Accuracy	0-shot	60.1	64.9
MBPP	pass@1	3-shot	56.6	63.6
HumanEval	pass@1	0-shot	66.5	75.0
LiveCodeBench	pass@1	0-shot	13.2	13.2
HiddenMath	Accuracy	0-shot	27.7	37.7
Global-MMLU-Lite	Accuracy	0-shot	59.0	64.5
MMLU (Pro)	Accuracy	0-shot	40.5	50.6

Android Performance Benchmarks with Google AI Edge

Note that all benchmark stats are from a Samsung S25 Ultra with 4096 KV cache size, 1024 tokens prefill, 256 tokens decode.

These numbers will continue to improve while Gemma 3n is in preview.

Weight Quantization	Backend	Prefill (tokens/sec)	Decode (tokens/sec)	Time to first token (sec)	Model size (MB)	Peak RSS Memory (MB)	GPU Memory (MB)
dynamic_int4	CPU	118	12.8	9.2	4201	3924	193
dynamic_int4	GPU	446	16.1	15.1	4201	5504	3048

Model size: measured by the size of the .tflite flatbuffer (serialization format for LiteRT models)
The inference on CPU is accelerated via the LiteRT XNNPACK delegate with 4 threads
Benchmark on CPU is done assuming XNNPACK cache is enabled
Benchmark on GPU is done assuming model is cached
Vision encoder is always run on GPU with 512×512 resolution
Cpufreq governor is set to performance during benchmark. Observed performance may vary depending on your phone’s hardware and current activity level.
dynamic_int4: quantized model with int4 weights and float activations.

0

Post5

By cspoweruser on

June 14, 2025

- Uncategorized

Gemma 3n model card

Model Page: Gemma 3n

Resources and Technical Documentation:

Responsible Generative AI Toolkit
Gemma on Kaggle
Google AI Edge documentation to run on mobile
Try on Android by downloading our Google AI Edge Gallery sample app

Terms of Use: Terms
Authors: Google DeepMind

Model Information

Summary description and brief definition of inputs and outputs.

Description

Gemma is a family of lightweight, state-of-the-art open models from Google, built from the same research and technology used to create the Gemini models. Gemma models are well-suited for a variety of content understanding tasks, including question answering, summarization, and reasoning. Their relatively small size makes it possible to deploy them in environments with limited resources such as laptops, desktops or your own cloud infrastructure, democratizing access to state of the art AI models and helping foster innovation for everyone.

Gemma 3n models are designed for efficient execution on low-resource devices. They are capable of multimodal input, handling text, image, video, and audio input, and generating text outputs, with open weights for instruction-tuned variants. These models were trained with data in over 140 spoken languages.

Gemma 3n models use selective parameter activation technology to reduce resource requirements. This technique allows the models to operate at an effective size of 2B and 4B parameters, which is lower than the total number of parameters they contain. For more information on Gemma 3n’s efficient parameter management technology, see the Gemma 3n page.

Inputs and outputs

Input:
- Text string, such as a question, a prompt, or a document to be summarized
- Images, normalized to 256×256, 512×512, or 768×768 resolution and encoded to 256 tokens each
- Audio data encoded to 6.25 tokens per second from a single channel
- Total input context of 32K tokens
Output:
- Generated text in response to the input, such as an answer to a question, analysis of image content, or a summary of a document
- Total output length up to 32K tokens, subtracting the request input tokens

Citation

@article{gemma_3n_2025,
    title={Gemma 3n},
    url={https://ai.google.dev/gemma/docs/gemma-3n},
    publisher={Google DeepMind},
    author={Gemma Team},
    year={2025}
}

Model Data

Data used for model training and how the data was processed.

Training Dataset

These models were trained on a dataset that includes a wide variety of sources totalling approximately 11 trillion tokens. The knowledge cutoff date for the training data was June 2024. Here are the key components:

Web Documents: A diverse collection of web text ensures the model is exposed to a broad range of linguistic styles, topics, and vocabulary. The training dataset includes content in over 140 languages.
Code: Exposing the model to code helps it to learn the syntax and patterns of programming languages, which improves its ability to generate code and understand code-related questions.
Mathematics: Training on mathematical text helps the model learn logical reasoning, symbolic representation, and to address mathematical queries.
Images: A wide range of images enables the model to perform image analysis and visual data extraction tasks.
Audio: A diverse set of sound samples enables the model to recognize speech, transcribe text from recordings, and identify information in audio data.

The combination of these diverse data sources is crucial for training a powerful multimodal model that can handle a wide variety of different tasks and data formats.

Data Preprocessing

Here are the key data cleaning and filtering methods applied to the training data:

CSAM Filtering: Rigorous CSAM (Child Sexual Abuse Material) filtering was applied at multiple stages in the data preparation process to ensure the exclusion of harmful and illegal content.
Sensitive Data Filtering: As part of making Gemma pre-trained models safe and reliable, automated techniques were used to filter out certain personal information and other sensitive data from training sets.
Additional methods: Filtering based on content quality and safety in line with our policies.

Implementation Information

Details about the model internals.

Hardware

Gemma was trained using Tensor Processing Unit (TPU) hardware (TPUv4p, TPUv5p and TPUv5e). Training generative models requires significant computational power. TPUs, designed specifically for matrix operations common in machine learning, offer several advantages in this domain:

Performance: TPUs are specifically designed to handle the massive computations involved in training generative models. They can speed up training considerably compared to CPUs.
Memory: TPUs often come with large amounts of high-bandwidth memory, allowing for the handling of large models and batch sizes during training. This can lead to better model quality.
Scalability: TPU Pods (large clusters of TPUs) provide a scalable solution for handling the growing complexity of large foundation models. You can distribute training across multiple TPU devices for faster and more efficient processing.
Cost-effectiveness: In many scenarios, TPUs can provide a more cost-effective solution for training large models compared to CPU-based infrastructure, especially when considering the time and resources saved due to faster training.

These advantages are aligned with Google’s commitments to operate sustainably.

Software

Training was done using JAX and ML Pathways. JAX allows researchers to take advantage of the latest generation of hardware, including TPUs, for faster and more efficient training of large models. ML Pathways is Google’s latest effort to build artificially intelligent systems capable of generalizing across multiple tasks. This is specially suitable for foundation models, including large language models like these ones.

Together, JAX and ML Pathways are used as described in the paper about the Gemini family of models: “the ‘single controller’ programming model of Jax and Pathways allows a single Python process to orchestrate the entire training run, dramatically simplifying the development workflow.”

Evaluation

Model evaluation metrics and results.

Benchmark Results

These models were evaluated at full precision (float32) against a large collection of different datasets and metrics to cover different aspects of content generation. Evaluation results marked with IT are for instruction-tuned models. Evaluation results marked with PT are for pre-trained models.

Reasoning and factuality

Benchmark	Metric	n-shot	E2B PT	E4B PT
HellaSwag	Accuracy	10-shot	72.2	78.6
BoolQ	Accuracy	0-shot	76.4	81.6
PIQA	Accuracy	0-shot	78.9	81.0
SocialIQA	Accuracy	0-shot	48.8	50.0
TriviaQA	Accuracy	5-shot	60.8	70.2
Natural Questions	Accuracy	5-shot	15.5	20.9
ARC-c	Accuracy	25-shot	51.7	61.6
ARC-e	Accuracy	0-shot	75.8	81.6
WinoGrande	Accuracy	5-shot	66.8	71.7
BIG-Bench Hard	Accuracy	few-shot	44.3	52.9
DROP	Token F1 score	1-shot	53.9	60.8

Multilingual

Benchmark	Metric	n-shot	E2B IT	E4B IT
MGSM	Accuracy	0-shot	53.1	60.7
WMT24++ (ChrF)	Character-level F-score	0-shot	42.7	50.1
Include	Accuracy	0-shot	38.6	57.2
MMLU (ProX)	Accuracy	0-shot	8.1	19.9
OpenAI MMLU	Accuracy	0-shot	22.3	35.6
Global-MMLU	Accuracy	0-shot	55.1	60.3
ECLeKTic	ECLeKTic score	0-shot	2.5	1.9

STEM and code

Benchmark	Metric	n-shot	E2B IT	E4B IT
GPQA Diamond	RelaxedAccuracy/accuracy	0-shot	24.8	23.7
LiveCodeBench v5	pass@1	0-shot	18.6	25.7
Codegolf v2.2	pass@1	0-shot	11.0	16.8
AIME 2025	Accuracy	0-shot	6.7	11.6

Additional benchmarks

Benchmark	Metric	n-shot	E2B IT	E4B IT
MMLU	Accuracy	0-shot	60.1	64.9
MBPP	pass@1	3-shot	56.6	63.6
HumanEval	pass@1	0-shot	66.5	75.0
LiveCodeBench	pass@1	0-shot	13.2	13.2
HiddenMath	Accuracy	0-shot	27.7	37.7
Global-MMLU-Lite	Accuracy	0-shot	59.0	64.5
MMLU (Pro)	Accuracy	0-shot	40.5	50.6

Android Performance Benchmarks with Google AI Edge

Note that all benchmark stats are from a Samsung S25 Ultra with 4096 KV cache size, 1024 tokens prefill, 256 tokens decode.

These numbers will continue to improve while Gemma 3n is in preview.

Weight Quantization	Backend	Prefill (tokens/sec)	Decode (tokens/sec)	Time to first token (sec)	Model size (MB)	Peak RSS Memory (MB)	GPU Memory (MB)
dynamic_int4	CPU	118	12.8	9.2	4201	3924	193
dynamic_int4	GPU	446	16.1	15.1	4201	5504	3048

Model size: measured by the size of the .tflite flatbuffer (serialization format for LiteRT models)
The inference on CPU is accelerated via the LiteRT XNNPACK delegate with 4 threads
Benchmark on CPU is done assuming XNNPACK cache is enabled
Benchmark on GPU is done assuming model is cached
Vision encoder is always run on GPU with 512×512 resolution
Cpufreq governor is set to performance during benchmark. Observed performance may vary depending on your phone’s hardware and current activity level.
dynamic_int4: quantized model with int4 weights and float activations.

0

Post4

By cspoweruser on

June 14, 2025

- Uncategorized

Gemma 3n model card

Model Page: Gemma 3n

Resources and Technical Documentation:

Responsible Generative AI Toolkit
Gemma on Kaggle
Google AI Edge documentation to run on mobile
Try on Android by downloading our Google AI Edge Gallery sample app

Terms of Use: Terms
Authors: Google DeepMind

Model Information

Summary description and brief definition of inputs and outputs.

Description

Gemma is a family of lightweight, state-of-the-art open models from Google, built from the same research and technology used to create the Gemini models. Gemma models are well-suited for a variety of content understanding tasks, including question answering, summarization, and reasoning. Their relatively small size makes it possible to deploy them in environments with limited resources such as laptops, desktops or your own cloud infrastructure, democratizing access to state of the art AI models and helping foster innovation for everyone.

Gemma 3n models are designed for efficient execution on low-resource devices. They are capable of multimodal input, handling text, image, video, and audio input, and generating text outputs, with open weights for instruction-tuned variants. These models were trained with data in over 140 spoken languages.

Gemma 3n models use selective parameter activation technology to reduce resource requirements. This technique allows the models to operate at an effective size of 2B and 4B parameters, which is lower than the total number of parameters they contain. For more information on Gemma 3n’s efficient parameter management technology, see the Gemma 3n page.

Inputs and outputs

Input:
- Text string, such as a question, a prompt, or a document to be summarized
- Images, normalized to 256×256, 512×512, or 768×768 resolution and encoded to 256 tokens each
- Audio data encoded to 6.25 tokens per second from a single channel
- Total input context of 32K tokens
Output:
- Generated text in response to the input, such as an answer to a question, analysis of image content, or a summary of a document
- Total output length up to 32K tokens, subtracting the request input tokens

Citation

@article{gemma_3n_2025,
    title={Gemma 3n},
    url={https://ai.google.dev/gemma/docs/gemma-3n},
    publisher={Google DeepMind},
    author={Gemma Team},
    year={2025}
}

Model Data

Data used for model training and how the data was processed.

Training Dataset

These models were trained on a dataset that includes a wide variety of sources totalling approximately 11 trillion tokens. The knowledge cutoff date for the training data was June 2024. Here are the key components:

Web Documents: A diverse collection of web text ensures the model is exposed to a broad range of linguistic styles, topics, and vocabulary. The training dataset includes content in over 140 languages.
Code: Exposing the model to code helps it to learn the syntax and patterns of programming languages, which improves its ability to generate code and understand code-related questions.
Mathematics: Training on mathematical text helps the model learn logical reasoning, symbolic representation, and to address mathematical queries.
Images: A wide range of images enables the model to perform image analysis and visual data extraction tasks.
Audio: A diverse set of sound samples enables the model to recognize speech, transcribe text from recordings, and identify information in audio data.

The combination of these diverse data sources is crucial for training a powerful multimodal model that can handle a wide variety of different tasks and data formats.

Data Preprocessing

Here are the key data cleaning and filtering methods applied to the training data:

CSAM Filtering: Rigorous CSAM (Child Sexual Abuse Material) filtering was applied at multiple stages in the data preparation process to ensure the exclusion of harmful and illegal content.
Sensitive Data Filtering: As part of making Gemma pre-trained models safe and reliable, automated techniques were used to filter out certain personal information and other sensitive data from training sets.
Additional methods: Filtering based on content quality and safety in line with our policies.

Implementation Information

Details about the model internals.

Hardware

Gemma was trained using Tensor Processing Unit (TPU) hardware (TPUv4p, TPUv5p and TPUv5e). Training generative models requires significant computational power. TPUs, designed specifically for matrix operations common in machine learning, offer several advantages in this domain:

Performance: TPUs are specifically designed to handle the massive computations involved in training generative models. They can speed up training considerably compared to CPUs.
Memory: TPUs often come with large amounts of high-bandwidth memory, allowing for the handling of large models and batch sizes during training. This can lead to better model quality.
Scalability: TPU Pods (large clusters of TPUs) provide a scalable solution for handling the growing complexity of large foundation models. You can distribute training across multiple TPU devices for faster and more efficient processing.
Cost-effectiveness: In many scenarios, TPUs can provide a more cost-effective solution for training large models compared to CPU-based infrastructure, especially when considering the time and resources saved due to faster training.

These advantages are aligned with Google’s commitments to operate sustainably.

Software

Training was done using JAX and ML Pathways. JAX allows researchers to take advantage of the latest generation of hardware, including TPUs, for faster and more efficient training of large models. ML Pathways is Google’s latest effort to build artificially intelligent systems capable of generalizing across multiple tasks. This is specially suitable for foundation models, including large language models like these ones.

Together, JAX and ML Pathways are used as described in the paper about the Gemini family of models: “the ‘single controller’ programming model of Jax and Pathways allows a single Python process to orchestrate the entire training run, dramatically simplifying the development workflow.”

Evaluation

Model evaluation metrics and results.

Benchmark Results

These models were evaluated at full precision (float32) against a large collection of different datasets and metrics to cover different aspects of content generation. Evaluation results marked with IT are for instruction-tuned models. Evaluation results marked with PT are for pre-trained models.

Reasoning and factuality

Benchmark	Metric	n-shot	E2B PT	E4B PT
HellaSwag	Accuracy	10-shot	72.2	78.6
BoolQ	Accuracy	0-shot	76.4	81.6
PIQA	Accuracy	0-shot	78.9	81.0
SocialIQA	Accuracy	0-shot	48.8	50.0
TriviaQA	Accuracy	5-shot	60.8	70.2
Natural Questions	Accuracy	5-shot	15.5	20.9
ARC-c	Accuracy	25-shot	51.7	61.6
ARC-e	Accuracy	0-shot	75.8	81.6
WinoGrande	Accuracy	5-shot	66.8	71.7
BIG-Bench Hard	Accuracy	few-shot	44.3	52.9
DROP	Token F1 score	1-shot	53.9	60.8

Multilingual

Benchmark	Metric	n-shot	E2B IT	E4B IT
MGSM	Accuracy	0-shot	53.1	60.7
WMT24++ (ChrF)	Character-level F-score	0-shot	42.7	50.1
Include	Accuracy	0-shot	38.6	57.2
MMLU (ProX)	Accuracy	0-shot	8.1	19.9
OpenAI MMLU	Accuracy	0-shot	22.3	35.6
Global-MMLU	Accuracy	0-shot	55.1	60.3
ECLeKTic	ECLeKTic score	0-shot	2.5	1.9

STEM and code

Benchmark	Metric	n-shot	E2B IT	E4B IT
GPQA Diamond	RelaxedAccuracy/accuracy	0-shot	24.8	23.7
LiveCodeBench v5	pass@1	0-shot	18.6	25.7
Codegolf v2.2	pass@1	0-shot	11.0	16.8
AIME 2025	Accuracy	0-shot	6.7	11.6

Additional benchmarks

Benchmark	Metric	n-shot	E2B IT	E4B IT
MMLU	Accuracy	0-shot	60.1	64.9
MBPP	pass@1	3-shot	56.6	63.6
HumanEval	pass@1	0-shot	66.5	75.0
LiveCodeBench	pass@1	0-shot	13.2	13.2
HiddenMath	Accuracy	0-shot	27.7	37.7
Global-MMLU-Lite	Accuracy	0-shot	59.0	64.5
MMLU (Pro)	Accuracy	0-shot	40.5	50.6

Android Performance Benchmarks with Google AI Edge

Note that all benchmark stats are from a Samsung S25 Ultra with 4096 KV cache size, 1024 tokens prefill, 256 tokens decode.

These numbers will continue to improve while Gemma 3n is in preview.

Weight Quantization	Backend	Prefill (tokens/sec)	Decode (tokens/sec)	Time to first token (sec)	Model size (MB)	Peak RSS Memory (MB)	GPU Memory (MB)
dynamic_int4	CPU	118	12.8	9.2	4201	3924	193
dynamic_int4	GPU	446	16.1	15.1	4201	5504	3048

Model size: measured by the size of the .tflite flatbuffer (serialization format for LiteRT models)
The inference on CPU is accelerated via the LiteRT XNNPACK delegate with 4 threads
Benchmark on CPU is done assuming XNNPACK cache is enabled
Benchmark on GPU is done assuming model is cached
Vision encoder is always run on GPU with 512×512 resolution
Cpufreq governor is set to performance during benchmark. Observed performance may vary depending on your phone’s hardware and current activity level.
dynamic_int4: quantized model with int4 weights and float activations.

0

Post3

By cspoweruser on

June 14, 2025

- Uncategorized

Gemma 3n model card

Model Page: Gemma 3n

Resources and Technical Documentation:

Responsible Generative AI Toolkit
Gemma on Kaggle
Google AI Edge documentation to run on mobile
Try on Android by downloading our Google AI Edge Gallery sample app

Terms of Use: Terms
Authors: Google DeepMind

Model Information

Summary description and brief definition of inputs and outputs.

Description

Gemma is a family of lightweight, state-of-the-art open models from Google, built from the same research and technology used to create the Gemini models. Gemma models are well-suited for a variety of content understanding tasks, including question answering, summarization, and reasoning. Their relatively small size makes it possible to deploy them in environments with limited resources such as laptops, desktops or your own cloud infrastructure, democratizing access to state of the art AI models and helping foster innovation for everyone.

Gemma 3n models are designed for efficient execution on low-resource devices. They are capable of multimodal input, handling text, image, video, and audio input, and generating text outputs, with open weights for instruction-tuned variants. These models were trained with data in over 140 spoken languages.

Gemma 3n models use selective parameter activation technology to reduce resource requirements. This technique allows the models to operate at an effective size of 2B and 4B parameters, which is lower than the total number of parameters they contain. For more information on Gemma 3n’s efficient parameter management technology, see the Gemma 3n page.

Inputs and outputs

Input:
- Text string, such as a question, a prompt, or a document to be summarized
- Images, normalized to 256×256, 512×512, or 768×768 resolution and encoded to 256 tokens each
- Audio data encoded to 6.25 tokens per second from a single channel
- Total input context of 32K tokens
Output:
- Generated text in response to the input, such as an answer to a question, analysis of image content, or a summary of a document
- Total output length up to 32K tokens, subtracting the request input tokens

Citation

@article{gemma_3n_2025,
    title={Gemma 3n},
    url={https://ai.google.dev/gemma/docs/gemma-3n},
    publisher={Google DeepMind},
    author={Gemma Team},
    year={2025}
}

Model Data

Data used for model training and how the data was processed.

Training Dataset

These models were trained on a dataset that includes a wide variety of sources totalling approximately 11 trillion tokens. The knowledge cutoff date for the training data was June 2024. Here are the key components:

Web Documents: A diverse collection of web text ensures the model is exposed to a broad range of linguistic styles, topics, and vocabulary. The training dataset includes content in over 140 languages.
Code: Exposing the model to code helps it to learn the syntax and patterns of programming languages, which improves its ability to generate code and understand code-related questions.
Mathematics: Training on mathematical text helps the model learn logical reasoning, symbolic representation, and to address mathematical queries.
Images: A wide range of images enables the model to perform image analysis and visual data extraction tasks.
Audio: A diverse set of sound samples enables the model to recognize speech, transcribe text from recordings, and identify information in audio data.

The combination of these diverse data sources is crucial for training a powerful multimodal model that can handle a wide variety of different tasks and data formats.

Data Preprocessing

Here are the key data cleaning and filtering methods applied to the training data:

CSAM Filtering: Rigorous CSAM (Child Sexual Abuse Material) filtering was applied at multiple stages in the data preparation process to ensure the exclusion of harmful and illegal content.
Sensitive Data Filtering: As part of making Gemma pre-trained models safe and reliable, automated techniques were used to filter out certain personal information and other sensitive data from training sets.
Additional methods: Filtering based on content quality and safety in line with our policies.

Implementation Information

Details about the model internals.

Hardware

Gemma was trained using Tensor Processing Unit (TPU) hardware (TPUv4p, TPUv5p and TPUv5e). Training generative models requires significant computational power. TPUs, designed specifically for matrix operations common in machine learning, offer several advantages in this domain:

Performance: TPUs are specifically designed to handle the massive computations involved in training generative models. They can speed up training considerably compared to CPUs.
Memory: TPUs often come with large amounts of high-bandwidth memory, allowing for the handling of large models and batch sizes during training. This can lead to better model quality.
Scalability: TPU Pods (large clusters of TPUs) provide a scalable solution for handling the growing complexity of large foundation models. You can distribute training across multiple TPU devices for faster and more efficient processing.
Cost-effectiveness: In many scenarios, TPUs can provide a more cost-effective solution for training large models compared to CPU-based infrastructure, especially when considering the time and resources saved due to faster training.

These advantages are aligned with Google’s commitments to operate sustainably.

Software

Training was done using JAX and ML Pathways. JAX allows researchers to take advantage of the latest generation of hardware, including TPUs, for faster and more efficient training of large models. ML Pathways is Google’s latest effort to build artificially intelligent systems capable of generalizing across multiple tasks. This is specially suitable for foundation models, including large language models like these ones.

Together, JAX and ML Pathways are used as described in the paper about the Gemini family of models: “the ‘single controller’ programming model of Jax and Pathways allows a single Python process to orchestrate the entire training run, dramatically simplifying the development workflow.”

Evaluation

Model evaluation metrics and results.

Benchmark Results

These models were evaluated at full precision (float32) against a large collection of different datasets and metrics to cover different aspects of content generation. Evaluation results marked with IT are for instruction-tuned models. Evaluation results marked with PT are for pre-trained models.

Reasoning and factuality

Benchmark	Metric	n-shot	E2B PT	E4B PT
HellaSwag	Accuracy	10-shot	72.2	78.6
BoolQ	Accuracy	0-shot	76.4	81.6
PIQA	Accuracy	0-shot	78.9	81.0
SocialIQA	Accuracy	0-shot	48.8	50.0
TriviaQA	Accuracy	5-shot	60.8	70.2
Natural Questions	Accuracy	5-shot	15.5	20.9
ARC-c	Accuracy	25-shot	51.7	61.6
ARC-e	Accuracy	0-shot	75.8	81.6
WinoGrande	Accuracy	5-shot	66.8	71.7
BIG-Bench Hard	Accuracy	few-shot	44.3	52.9
DROP	Token F1 score	1-shot	53.9	60.8

Multilingual

Benchmark	Metric	n-shot	E2B IT	E4B IT
MGSM	Accuracy	0-shot	53.1	60.7
WMT24++ (ChrF)	Character-level F-score	0-shot	42.7	50.1
Include	Accuracy	0-shot	38.6	57.2
MMLU (ProX)	Accuracy	0-shot	8.1	19.9
OpenAI MMLU	Accuracy	0-shot	22.3	35.6
Global-MMLU	Accuracy	0-shot	55.1	60.3
ECLeKTic	ECLeKTic score	0-shot	2.5	1.9

STEM and code

Benchmark	Metric	n-shot	E2B IT	E4B IT
GPQA Diamond	RelaxedAccuracy/accuracy	0-shot	24.8	23.7
LiveCodeBench v5	pass@1	0-shot	18.6	25.7
Codegolf v2.2	pass@1	0-shot	11.0	16.8
AIME 2025	Accuracy	0-shot	6.7	11.6

Additional benchmarks

Benchmark	Metric	n-shot	E2B IT	E4B IT
MMLU	Accuracy	0-shot	60.1	64.9
MBPP	pass@1	3-shot	56.6	63.6
HumanEval	pass@1	0-shot	66.5	75.0
LiveCodeBench	pass@1	0-shot	13.2	13.2
HiddenMath	Accuracy	0-shot	27.7	37.7
Global-MMLU-Lite	Accuracy	0-shot	59.0	64.5
MMLU (Pro)	Accuracy	0-shot	40.5	50.6

Android Performance Benchmarks with Google AI Edge

Note that all benchmark stats are from a Samsung S25 Ultra with 4096 KV cache size, 1024 tokens prefill, 256 tokens decode.

These numbers will continue to improve while Gemma 3n is in preview.

Weight Quantization	Backend	Prefill (tokens/sec)	Decode (tokens/sec)	Time to first token (sec)	Model size (MB)	Peak RSS Memory (MB)	GPU Memory (MB)
dynamic_int4	CPU	118	12.8	9.2	4201	3924	193
dynamic_int4	GPU	446	16.1	15.1	4201	5504	3048

Model size: measured by the size of the .tflite flatbuffer (serialization format for LiteRT models)
The inference on CPU is accelerated via the LiteRT XNNPACK delegate with 4 threads
Benchmark on CPU is done assuming XNNPACK cache is enabled
Benchmark on GPU is done assuming model is cached
Vision encoder is always run on GPU with 512×512 resolution
Cpufreq governor is set to performance during benchmark. Observed performance may vary depending on your phone’s hardware and current activity level.
dynamic_int4: quantized model with int4 weights and float activations.

0

Post2

By cspoweruser on

June 14, 2025

- Uncategorized

Gemma 3n model card

Model Page: Gemma 3n

Resources and Technical Documentation:

Responsible Generative AI Toolkit
Gemma on Kaggle
Google AI Edge documentation to run on mobile
Try on Android by downloading our Google AI Edge Gallery sample app

Terms of Use: Terms
Authors: Google DeepMind

Model Information

Summary description and brief definition of inputs and outputs.

Description

Gemma is a family of lightweight, state-of-the-art open models from Google, built from the same research and technology used to create the Gemini models. Gemma models are well-suited for a variety of content understanding tasks, including question answering, summarization, and reasoning. Their relatively small size makes it possible to deploy them in environments with limited resources such as laptops, desktops or your own cloud infrastructure, democratizing access to state of the art AI models and helping foster innovation for everyone.

Gemma 3n models are designed for efficient execution on low-resource devices. They are capable of multimodal input, handling text, image, video, and audio input, and generating text outputs, with open weights for instruction-tuned variants. These models were trained with data in over 140 spoken languages.

Gemma 3n models use selective parameter activation technology to reduce resource requirements. This technique allows the models to operate at an effective size of 2B and 4B parameters, which is lower than the total number of parameters they contain. For more information on Gemma 3n’s efficient parameter management technology, see the Gemma 3n page.

Inputs and outputs

Input:
- Text string, such as a question, a prompt, or a document to be summarized
- Images, normalized to 256×256, 512×512, or 768×768 resolution and encoded to 256 tokens each
- Audio data encoded to 6.25 tokens per second from a single channel
- Total input context of 32K tokens
Output:
- Generated text in response to the input, such as an answer to a question, analysis of image content, or a summary of a document
- Total output length up to 32K tokens, subtracting the request input tokens

Citation

@article{gemma_3n_2025,
    title={Gemma 3n},
    url={https://ai.google.dev/gemma/docs/gemma-3n},
    publisher={Google DeepMind},
    author={Gemma Team},
    year={2025}
}

Model Data

Data used for model training and how the data was processed.

Training Dataset

These models were trained on a dataset that includes a wide variety of sources totalling approximately 11 trillion tokens. The knowledge cutoff date for the training data was June 2024. Here are the key components:

Web Documents: A diverse collection of web text ensures the model is exposed to a broad range of linguistic styles, topics, and vocabulary. The training dataset includes content in over 140 languages.
Code: Exposing the model to code helps it to learn the syntax and patterns of programming languages, which improves its ability to generate code and understand code-related questions.
Mathematics: Training on mathematical text helps the model learn logical reasoning, symbolic representation, and to address mathematical queries.
Images: A wide range of images enables the model to perform image analysis and visual data extraction tasks.
Audio: A diverse set of sound samples enables the model to recognize speech, transcribe text from recordings, and identify information in audio data.

The combination of these diverse data sources is crucial for training a powerful multimodal model that can handle a wide variety of different tasks and data formats.

Data Preprocessing

Here are the key data cleaning and filtering methods applied to the training data:

CSAM Filtering: Rigorous CSAM (Child Sexual Abuse Material) filtering was applied at multiple stages in the data preparation process to ensure the exclusion of harmful and illegal content.
Sensitive Data Filtering: As part of making Gemma pre-trained models safe and reliable, automated techniques were used to filter out certain personal information and other sensitive data from training sets.
Additional methods: Filtering based on content quality and safety in line with our policies.

Implementation Information

Details about the model internals.

Hardware

Gemma was trained using Tensor Processing Unit (TPU) hardware (TPUv4p, TPUv5p and TPUv5e). Training generative models requires significant computational power. TPUs, designed specifically for matrix operations common in machine learning, offer several advantages in this domain:

Performance: TPUs are specifically designed to handle the massive computations involved in training generative models. They can speed up training considerably compared to CPUs.
Memory: TPUs often come with large amounts of high-bandwidth memory, allowing for the handling of large models and batch sizes during training. This can lead to better model quality.
Scalability: TPU Pods (large clusters of TPUs) provide a scalable solution for handling the growing complexity of large foundation models. You can distribute training across multiple TPU devices for faster and more efficient processing.
Cost-effectiveness: In many scenarios, TPUs can provide a more cost-effective solution for training large models compared to CPU-based infrastructure, especially when considering the time and resources saved due to faster training.

These advantages are aligned with Google’s commitments to operate sustainably.

Software

Training was done using JAX and ML Pathways. JAX allows researchers to take advantage of the latest generation of hardware, including TPUs, for faster and more efficient training of large models. ML Pathways is Google’s latest effort to build artificially intelligent systems capable of generalizing across multiple tasks. This is specially suitable for foundation models, including large language models like these ones.

Together, JAX and ML Pathways are used as described in the paper about the Gemini family of models: “the ‘single controller’ programming model of Jax and Pathways allows a single Python process to orchestrate the entire training run, dramatically simplifying the development workflow.”

Evaluation

Model evaluation metrics and results.

Benchmark Results

These models were evaluated at full precision (float32) against a large collection of different datasets and metrics to cover different aspects of content generation. Evaluation results marked with IT are for instruction-tuned models. Evaluation results marked with PT are for pre-trained models.

Reasoning and factuality

Benchmark	Metric	n-shot	E2B PT	E4B PT
HellaSwag	Accuracy	10-shot	72.2	78.6
BoolQ	Accuracy	0-shot	76.4	81.6
PIQA	Accuracy	0-shot	78.9	81.0
SocialIQA	Accuracy	0-shot	48.8	50.0
TriviaQA	Accuracy	5-shot	60.8	70.2
Natural Questions	Accuracy	5-shot	15.5	20.9
ARC-c	Accuracy	25-shot	51.7	61.6
ARC-e	Accuracy	0-shot	75.8	81.6
WinoGrande	Accuracy	5-shot	66.8	71.7
BIG-Bench Hard	Accuracy	few-shot	44.3	52.9
DROP	Token F1 score	1-shot	53.9	60.8

Multilingual

Benchmark	Metric	n-shot	E2B IT	E4B IT
MGSM	Accuracy	0-shot	53.1	60.7
WMT24++ (ChrF)	Character-level F-score	0-shot	42.7	50.1
Include	Accuracy	0-shot	38.6	57.2
MMLU (ProX)	Accuracy	0-shot	8.1	19.9
OpenAI MMLU	Accuracy	0-shot	22.3	35.6
Global-MMLU	Accuracy	0-shot	55.1	60.3
ECLeKTic	ECLeKTic score	0-shot	2.5	1.9

STEM and code

Benchmark	Metric	n-shot	E2B IT	E4B IT
GPQA Diamond	RelaxedAccuracy/accuracy	0-shot	24.8	23.7
LiveCodeBench v5	pass@1	0-shot	18.6	25.7
Codegolf v2.2	pass@1	0-shot	11.0	16.8
AIME 2025	Accuracy	0-shot	6.7	11.6

Additional benchmarks

Benchmark	Metric	n-shot	E2B IT	E4B IT
MMLU	Accuracy	0-shot	60.1	64.9
MBPP	pass@1	3-shot	56.6	63.6
HumanEval	pass@1	0-shot	66.5	75.0
LiveCodeBench	pass@1	0-shot	13.2	13.2
HiddenMath	Accuracy	0-shot	27.7	37.7
Global-MMLU-Lite	Accuracy	0-shot	59.0	64.5
MMLU (Pro)	Accuracy	0-shot	40.5	50.6

Android Performance Benchmarks with Google AI Edge

Note that all benchmark stats are from a Samsung S25 Ultra with 4096 KV cache size, 1024 tokens prefill, 256 tokens decode.

These numbers will continue to improve while Gemma 3n is in preview.

Weight Quantization	Backend	Prefill (tokens/sec)	Decode (tokens/sec)	Time to first token (sec)	Model size (MB)	Peak RSS Memory (MB)	GPU Memory (MB)
dynamic_int4	CPU	118	12.8	9.2	4201	3924	193
dynamic_int4	GPU	446	16.1	15.1	4201	5504	3048

Model size: measured by the size of the .tflite flatbuffer (serialization format for LiteRT models)
The inference on CPU is accelerated via the LiteRT XNNPACK delegate with 4 threads
Benchmark on CPU is done assuming XNNPACK cache is enabled
Benchmark on GPU is done assuming model is cached
Vision encoder is always run on GPU with 512×512 resolution
Cpufreq governor is set to performance during benchmark. Observed performance may vary depending on your phone’s hardware and current activity level.
dynamic_int4: quantized model with int4 weights and float activations.

0

Post1

By cspoweruser on

June 14, 2025

- Uncategorized

Model Card for Magistral-Small-2506

Building upon Mistral Small 3.1 (2503), with added reasoning capabilities, undergoing SFT from Magistral Medium traces and RL on top, it’s a small, efficient reasoning model with 24B parameters.

Magistral Small can be deployed locally, fitting within a single RTX 4090 or a 32GB RAM MacBook once quantized.

Learn more about Magistral in our blog post.

Key Features

Reasoning: Capable of long chains of reasoning traces before providing an answer.
Multilingual: Supports dozens of languages, including English, French, German, Greek, Hindi, Indonesian, Italian, Japanese, Korean, Malay, Nepali, Polish, Portuguese, Romanian, Russian, Serbian, Spanish, Swedish, Turkish, Ukrainian, Vietnamese, Arabic, Bengali, Chinese, and Farsi.
Apache 2.0 License: Open license allowing usage and modification for both commercial and non-commercial purposes.
Context Window: A 128k context window, but performance might degrade past 40k. Hence we recommend setting the maximum model length to 40k.

Benchmark Results

Model	AIME24 pass@1	AIME25 pass@1	GPQA Diamond	Livecodebench (v5)
Magistral Medium	73.59%	64.95%	70.83%	59.36%
Magistral Small	70.68%	62.76%	68.18%	55.84%

Sampling parameters

Please make sure to use:

top_p: 0.95
temperature: 0.7
max_tokens: 40960

Basic Chat Template

We highly recommend including the default system prompt used during RL for the best results, you can edit and customise it if needed for your specific use case.

<s>[SYSTEM_PROMPT]system_prompt

A user will ask you to solve a task. You should first draft your thinking process (inner monologue) until you have derived the final answer. Afterwards, write a self-contained summary of your thoughts (i.e. your summary should be succinct but contain all the critical steps you needed to reach the conclusion). You should use Markdown to format your response. Write both your thoughts and summary in the same language as the task posed by the user. NEVER use \boxed{} in your response.

Your thinking process must follow the template below:
<think>
Your thoughts or/and draft, like working through an exercise on scratch paper. Be as casual and as long as you want until you are confident to generate a correct answer.
</think>

Here, provide a concise summary that reflects your reasoning and presents a clear final answer to the user. Don't mention that this is a summary.

Problem:

[/SYSTEM_PROMPT][INST]user_message[/INST]<think>
reasoning_traces
</think>
assistant_response</s>[INST]user_message[/INST]

system_prompt, user_message and assistant_response are placeholders.

We invite you to choose, depending on your use case and requirements, between keeping reasoning traces during multi-turn interactions or keeping only the final assistant response.

Please make sure to use mistral-common as the source of truth

Usage

The model can be used with the following frameworks;

Inference

vllm (recommended): See below

In addition the community has prepared quantized versions of the model that can be used with the following frameworks (alphabetically sorted):

Training

Fine-tuning is possible with (alphabetically sorted):

Other

Also you can use Magistral with:

kaggle: https://www.kaggle.com/models/mistral-ai/magistral-small-2506

vLLM (recommended)

We recommend using this model with the vLLM library to implement production-ready inference pipelines.

Installation

Make sure you install the latest vLLM code:

pip install -U vllm \
    --pre \
    --extra-index-url https://wheels.vllm.ai/nightly

Doing so should automatically install mistral_common >= 1.6.0.

0