News
ResearchOpenAI pens AI safety agreement with four firmsReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierOpenAI鈥檚 New Hacking Model Found Dozens Of Holes In Apple TechFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIResearchOpenAI pens AI safety agreement with four firmsReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierOpenAI鈥檚 New Hacking Model Found Dozens Of Holes In Apple TechFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAI
EveryLLM
fnMinistral 3B and 8B

Ministral 3B and 8B

Two small Mistral AI models built for on-device and edge computing.

Ministral 3B and 8B are a pair of small language models from Mistral AI, released in October 2024 on the first anniversary of Mistral 7B. The company positions them for private, low-latency work on local devices and at the edge, including translation, offline assistants, analytics and robotics.

Share on WhatsApp

  • Released October 2024, one year after Mistral 7B
  • Two sizes: 3B and 8B parameters
  • Context windows up to 128k tokens, 32k on vLLM
  • 8B Instruct weights downloadable for research use
  • API pricing from $0.04 per million tokens (8B)

What it is

Ministral 3B and 8B are two compact language models announced by Mistral AI in October 2024. The company describes them as its smallest offerings, aimed at running locally rather than in a data centre. Sizes are 3 billion and 8 billion parameters. Intended jobs include on-device translation, assistants that work without internet access, local data analysis, autonomous robotics and the routing steps of multi-step agent systems. Mistral AI claims they lead the sub-10B class in knowledge, commonsense, reasoning, function calling and efficiency, and says its own re-evaluation framework was used for the comparisons it published.

How it came about

The pair arrived exactly one year after Mistral 7B, the model Mistral AI credits with opening up independent frontier AI work. The company says demand from customers and partners for privacy-first local inference drove the project. Both sizes shipped on the announcement day. Weights for the 8B Instruct version were released for research use at that point, with cloud partner availability promised shortly afterwards.

Inside the model

Both models handle windows of up to 128k tokens, though the company notes vLLM support currently caps this at 32k. The larger of the two uses an interleaved sliding-window attention scheme, which limits how much of the past context each step reads so inference runs faster and uses less memory. Mistral AI offers help with lossless quantization, a way of shrinking weights without losing accuracy, for customers deploying the models themselves. The models can also be fine-tuned for narrow tasks such as parsing inputs, routing work and issuing API calls.

What it can do

Mistral AI reports that both sizes beat comparable small open models across the categories it tested, using an internal evaluation framework it says gives a fair comparison. Charts cover base and Instruct versions separately. The company states that its 3B model already exceeds Mistral 7B on most benchmarks one year after that model appeared. No individual scores are given in the announcement text.

How to use it

Both models were offered through Mistral AI's platform from launch, priced at $0.1 per million tokens for the 3B and $0.04 per million tokens for the 8B, covering input and output. Self-deployment requires a commercial licence obtained from the company. The 8B Instruct weights are downloadable for research purposes. Distribution through cloud partners was announced as coming soon.

Safety and testing

The announcement does not describe safety evaluations, red-teaming or incidents for these models.

Reception and debate

The company's own announcement presents the models as leaders in their size class and highlights customer demand for local, private inference. Independent reception is not covered here.

Why it matters

Mistral AI frames the pair as evidence of how quickly small models have advanced, noting that its 3B offering surpasses the 7B model released only a year earlier on most benchmarks. The company suggests the models serve as cheap intermediaries alongside larger systems such as Mistral Large in agentic pipelines.

Questions people ask about Ministral 3B and 8B

What are Ministral 3B and 8B?

They are two small language models from Mistral AI, sized at 3 and 8 billion parameters. The company built them for running on local devices and at the edge, where privacy, low latency and low cost matter more than raw scale.

How do I run them?

Mistral AI offered both through its platform at launch, with per-token pricing. Running them on your own hardware needs a commercial licence from the company, which also offers help with quantization. Research users can download the 8B Instruct weights.

How much context can they handle?

Up to 128k tokens according to Mistral AI, though the company notes that vLLM support currently reaches only 32k. The 8B model uses a sliding-window attention pattern to keep inference fast and memory use low.

How do they compare with earlier Mistral models?

Mistral AI says its 3B model beats Mistral 7B on most benchmarks despite being less than half the size. Comparisons against other small open models were run with the company's own evaluation framework.

What are they used for?

The company points to on-device translation, assistants that function offline, local analytics, robotics, and the parsing and routing steps inside multi-step agent workflows alongside larger models such as Mistral Large.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Ministral 3B and 8B yet, so the facts on this page come from the sources listed above: mistral.ai/news/ministraux. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.