Citation and evidence

Mistral Small

15 min full readUpdated 25 references

This article's verification

Report a problem with this article

Edit article
More

Use this article

Raw MarkdownExplore connections

Improve this page

Revision historyDiscuss this article

Topics

AI ModelsLarge Language ModelsOpen Source AI

Cite this article

Mistral Small is a line of large language models from the French company Mistral AI, positioned between the company's smallest models and its flagship Mistral Large. The name first appeared in December 2023 as an API endpoint that served Mixtral 8x7B,[1] and from February 2024 it named a new Mistral model optimized for low-latency workloads.[2] Versions released since then include a 22-billion-parameter model under the non-commercial Mistral Research License (September 2024),[3][4] three dense 24-billion-parameter models under the Apache 2.0 license (Mistral Small 3, 3.1 and 3.2, released between January and June 2025),[6][9][12] and Mistral Small 4, a sparse mixture-of-experts model with 119 billion total parameters released in March 2026.[15] Several other Mistral models, among them Devstral Small, Magistral Small and Voxtral Small, were trained on top of Mistral Small checkpoints.[17][18][19][20]

History

The mistral-small endpoint (December 2023)

Mistral opened beta access to its developer platform, La Plateforme, on December 11, 2023, with three chat endpoints of increasing capability: mistral-tiny, mistral-small and mistral-medium.[1] The mistral-small endpoint served Mixtral 8x7B, which Mistral called its newest model at the time; the company said it handled English, French, Italian, German, Spanish and code, and reported an MT-Bench score of 8.3.[1] Mistral-tiny served Mistral 7B Instruct v0.2, and mistral-medium served a prototype model.[1]

Mistral Small 24.02 (February 2024)

On February 26, 2024, in the same announcement that introduced Mistral Large, Mistral described "a new Mistral Small, optimised for low latency workloads," available through the endpoint mistral-small-2402.[2] The company said it outperformed Mixtral 8x7B with lower latency; the announcement itself gave no figures for it, pointing readers to benchmarks in Mistral's documentation instead, and presented it as an intermediate option between its open-weight models and its flagship.[2] Mistral Small shared Mistral Large's support for retrieval-augmented generation and function calling, and at launch function calling and a JSON output mode were available only on mistral-small and mistral-large.[2] Mistral's model documentation now lists this version as Mistral Small 1.0.[16]

Mistral Small 24.09 (September 2024)

Mistral released Mistral Small v24.09 on September 17, 2024, calling it an upgrade of v24.02 with improvements in human alignment, reasoning and code.[3] It had 22 billion parameters and was presented as a mid-point between the 12B Mistral NeMo and Mistral Large 2, aimed at tasks such as translation, summarization and sentiment analysis.[3] The weights were published on Hugging Face as Mistral-Small-Instruct-2409, with a 32k-token sequence length, a vocabulary of 32,768 tokens and function-calling support.[4] They were licensed under the Mistral Research License (MRL), which allowed self-deployment only for non-commercial purposes.[3]

The same announcement cut the API price of Mistral Small by 80 percent, from $1 to $0.20 per million input tokens and from $3 to $0.60 per million output tokens, as part of a price reduction across Mistral's lineup.[3] Mistral's documentation lists this version as Mistral Small 2.0.[16]

Mistral Small 3 (January 2025)

Mistral Small 3 was announced on January 30, 2025 as "a latency-optimized 24B-parameter model released under the Apache 2.0 license."[6] Both the pretrained base checkpoint and the instruction-tuned checkpoint were released under Apache 2.0, and the announcement said the company was "renewing our commitment to using Apache 2.0 license for our general purpose models, as we progressively move away from MRL-licensed models."[6] Mistral reported more than 81 percent accuracy on MMLU and a speed of 150 tokens per second, and said the model was competitive with larger models such as Llama 3.3 70B and Qwen 32B while running more than three times faster than Llama 3.3 70B on the same hardware.[6] It attributed the speed partly to the model having "far fewer layers than competing models."[6] The company also stated that the model was "neither trained with RL nor synthetic data," and suggested it as a base for building reasoning models.[6]

Mistral pitched the model as an open alternative to proprietary models such as GPT-4o mini, and listed uses including fast conversational assistance, low-latency function calling, fine-tuning for subject-matter experts and local inference: once quantized, it could run on a single RTX 4090 or a MacBook with 32 GB of RAM.[6] On launch day it was available on La Plateforme as mistral-small-2501 and through Hugging Face, Ollama, Kaggle, Together AI, Fireworks AI and IBM watsonx.[6] The Hugging Face model card lists a 32k context window, the Tekken tokenizer with a 131k vocabulary, and support for languages including English, French, German, Spanish, Italian, Chinese, Japanese, Korean, Portuguese, Dutch and Polish.[7]

Mistral Small 3.1 (March 2025)

Mistral Small 3.1, announced on March 17, 2025, kept the 24B size and the Apache 2.0 license and added image understanding and a context window of up to 128k tokens.[9][10] Mistral said it outperformed comparable models such as Gemma 3 and GPT-4o mini while running at 150 tokens per second, and again said it could run on a single RTX 4090 or a Mac with 32 GB of RAM.[9] Base and instruct checkpoints were released on Hugging Face, and the model was offered on La Plateforme and Google Vertex AI, with NVIDIA NIM and Azure AI Foundry announced for the following weeks.[9] The model card describes native function calling and JSON output and support for dozens of languages.[10]

Mistral Small 3.2 (June 2025)

Mistral Small 3.2 (Mistral-Small-3.2-24B-Instruct-2506) was a minor update to the 3.1 instruct model, built on the same 3.1 base checkpoint.[12] Mistral's documentation dates its release to June 20, 2025, under the API identifier mistral-small-2506, with a 128k context window and an Apache 2.0 license.[13] The model card lists three targeted changes: better instruction following, fewer infinite generations and repetitive answers, and a more robust function-calling template.[12] Mistral reported that its rate of infinite generations on an internal set of long, repetitive prompts fell from 2.11 percent to 1.29 percent, and that scores in other categories should "match or slightly improve" relative to 3.1.[12]

Mistral Small Creative (December 2025)

Mistral Small Creative was an experimental model offered through Mistral's Labs program from December 11, 2025, under the API identifier labs-mistral-small-creative.[14] The documentation describes it as "an experimental specialized small model trained on meticulously curated data," designed for creative writing, narrative generation, roleplay and character-driven dialogue, with a 32k context window.[14] The page does not state a parameter count or whether weights were released. It was deprecated on March 31, 2026, with Ministral 3 8B named as the model to use for new integrations.[14][16]

Mistral Small 4 (March 2026)

Mistral Small 4, released on March 16, 2026, broke with the dense design of the 3.x models. It is a mixture-of-experts model with 119 billion total parameters across 128 experts, of which 4 are active for each token, giving about 6 billion active parameters (8 billion including the embedding and output layers).[15] Mistral's API documentation gives the active count as 6.5 billion.[23] It has a 256k context window, an Apache 2.0 license and a reasoning_effort setting that switches between fast responses in the style of Mistral Small 3.2 and step-by-step reasoning in the style of Magistral.[15] Mistral described it as the first of its models to combine the capabilities of Magistral, Pixtral, Devstral and the Mistral Small instruct models, and reported a 40 percent cut in end-to-end completion time and three times the requests per second compared with Mistral Small 3.[15] Its minimum listed deployment hardware is four NVIDIA HGX H100, two HGX H200 or one DGX B200 system, well beyond the single consumer GPU that suited the 24B models.[15] The model is covered in detail in the Mistral Small 4 article.

Versions

VersionReleasedAPI identifierParametersContextWeights license
mistral-small endpoint (Mixtral 8x7B)2023-12-11mistral-smallMixtral 8x7Bnot statednot applicable (endpoint)[1]
Mistral Small 1.0 (24.02)2024-02-26mistral-small-2402not disclosednot statedAPI endpoint[2][16]
Mistral Small 2.0 (24.09)2024-09-17mistral-small-240922B dense32kMistral Research License[3][4]
Mistral Small 3 (25.01)2025-01-30mistral-small-250124B dense32kApache 2.0[6][7]
Mistral Small 3.1 (25.03)2025-03-17mistral-small-250324B dense, plus vision encoder128kApache 2.0[9][10]
Mistral Small 3.2 (25.06)2025-06-20mistral-small-250624B dense, plus vision encoder128kApache 2.0[12][13]
Mistral Small Creative (25.12)2025-12-11labs-mistral-small-creativenot stated32knot stated[14]
Mistral Small 4 (26.03)2026-03-16mistral-small-2603119B MoE, about 6B active256kApache 2.0[15][23]

Expanded article table

Mistral's documentation numbers the API versions 1.0 (24.02), 2.0 (24.09), 3.0 (25.01), 3.1 (25.03) and 3.2 (25.06).[16]

Architecture

The 22B and 24B Mistral Small models are dense decoder-only transformers. Their published configuration files show that Mistral Small 3 was shallower and wider in its feed-forward layers than the 22B model it replaced, consistent with the "far fewer layers" Mistral cited as a source of its speed.[5][8][6] Both use grouped-query attention, with 8 key-value heads shared across the query heads.[5][8]

SettingSmall 24.09 (22B)Small 3 (24B)Small 3.1 and 3.2 (24B)
Layers564040
Hidden size6,1445,1205,120
Feed-forward size16,38432,76832,768
Attention heads (query / key-value)48 / 832 / 832 / 8
Vocabulary32,768131,072131,072
Maximum positions32,76832,768131,072
Vision encodernonenone24-layer Pixtral-type encoder
Source[5][8][11][24]

Expanded article table

From version 3.1 the checkpoints add a vision encoder (24 layers, hidden size 1,024, 14-pixel patches) whose configuration is labelled as the Pixtral type.[11] The rotary position embedding base value (rope_theta) is 100,000,000 in Mistral Small 3 and 1,000,000,000 in 3.1.[8][11] Mistral's model cards say that running the 24B models unquantized on a GPU takes about 55 GB of memory in bf16 or fp16.[7][10][12]

Benchmarks

All figures below were reported by Mistral on its model cards, from its own evaluation pipeline, and have not been independently audited.

Mistral Small 3 instruct, compared with other models in the 2501 model card:[7]

BenchmarkMistral Small 3Gemma 2 27BLlama 3.3 70BQwen 2.5 32BGPT-4o mini
MMLU-Pro (5-shot CoT)0.6630.5360.6660.6830.617
GPQA Main (5-shot CoT)0.4530.3440.5310.4040.377
HumanEval (pass@1)0.8480.7320.8540.9090.890
MT-Bench (dev)8.357.867.968.268.33
Arena Hard0.8730.7880.8400.8600.897
IFEval0.8290.80650.88350.84010.8499

Expanded article table

Mistral Small 3.1 instruct, compared with other models in the 2503 model card:[10]

BenchmarkMistral Small 3.1Gemma 3 27BGPT-4o miniClaude 3.5 Haiku
MMLU80.62%76.90%82.00%77.60%
MMLU-Pro66.76%67.50%61.70%65.00%
MATH69.30%89.00%70.20%69.20%
GPQA Main44.42%36.83%40.20%37.05%
HumanEval88.41%87.80%87.20%88.10%
MMMU64.00%64.90%59.40%60.50%
MathVista68.91%67.60%56.70%61.60%
RULER 128k81.20%66.00%65.80%91.90%
LongBench v237.18%34.59%29.30%35.19%

Expanded article table

Mistral Small 3.2 instruct against 3.1 instruct, from the 2506 model card:[12]

BenchmarkSmall 3.1Small 3.2
WildBench v255.6%65.33%
Arena Hard v219.56%43.1%
Internal instruction-following accuracy82.75%84.78%
Infinite generations (lower is better)2.11%1.29%
MMLU-Pro (5-shot CoT)66.76%69.06%
HumanEval Plus (pass@5)88.99%92.90%
MBPP Plus (pass@5)74.63%78.33%
MMLU80.62%80.50%
GPQA Main (5-shot CoT)44.42%44.22%
MMMU64.00%62.50%

Expanded article table

The 3.2 table shows the update was not uniformly better: MMLU, GPQA Main, MMMU and some other vision scores dipped slightly while instruction following and coding improved.[12]

Licensing

The line's licensing changed twice. The 24.02 model was announced as an API endpoint, placed between Mistral's open-weight models and its flagship.[2] The 24.09 weights were downloadable but covered by the Mistral Research License, under which self-deployment was permitted for non-commercial purposes.[3] Developer Simon Willison noted that earlier Mistral Small models had used a license that ruled out commercial use without a separately negotiated agreement, and read the January 2025 announcement as a move away from that license for Mistral's core models.[21] Mistral Small 3, 3.1, 3.2 and 4 were all released under Apache 2.0, which permits commercial and non-commercial use and modification.[6][9][12][15] Willison also reported that Mistral Small 3's API price was half that of the previous Mistral Small API model, at $0.10 per million input tokens and $0.30 per million output tokens.[21]

Derived models

The 24B checkpoints served as the starting point for several later Mistral models:

ModelBuilt onSource
Devstral Small (2505), coding agentMistral Small 3.1, with the vision encoder removed[17]
Magistral Small (2506), reasoningMistral Small 3.1 (2503), with supervised fine-tuning on Magistral Medium traces and reinforcement learning[18]
Magistral Small 1.2 (2509), reasoningMistral Small 3.2 (2506)[19]
Voxtral Small (2507), audio inputMistral Small 3.1 24B backbone, according to the Voxtral paper (the model card links the Mistral Small 3 base checkpoint)[20][25]

Expanded article table

Outside Mistral, the Mistral Small 3.1 announcement pointed to DeepHermes 24B from Nous Research as a reasoning model built on Mistral Small 3.[9]

Deprecation and retirement

Mistral retires older API versions on a published schedule, and as of November 2026 its documentation lists every Mistral Small API version before 4 as deprecated or retired:[16]

API identifierDeprecationRetirementAlternative named by Mistral
mistral-small-24022024-11-302025-06-16Mistral Small 4
mistral-small-24092025-11-062025-11-30Mistral Small 4
mistral-small-25012025-11-062025-11-30Mistral Small 4
mistral-small-25032025-11-062025-11-30Mistral Small 4
mistral-small-25062026-04-302026-07-31Mistral Small 4
labs-mistral-small-creative2026-03-312026-04-30Ministral 3 8B

Expanded article table

The open-weight checkpoints of the 24.09, 3, 3.1 and 3.2 releases remain listed on Mistral's Hugging Face organization; the retirement dates apply to Mistral's hosted API.[4][7][10][12][16]

Reception

Coverage of Mistral Small 3 focused on its size-to-quality ratio and its license. SiliconANGLE noted that Llama 3.3 70B, the model Mistral compared it with, has more than three times as many parameters, and reported Mistral's description of the model as catering to "the '80%' of generative AI tasks."[22] Willison welcomed a new open model in the roughly 20 GB class that he regarded as a good balance between memory use and usefulness, ran the 14 GB Ollama build locally, and observed that despite its name it appeared to be "the fourth release of a model under that label."[21]

Continue reading

More to read

Topics:AI ModelsLarge Language ModelsOpen Source AI

References

  1. ^1 ^2 ^3 ^4 ^5Mistral AI. "La Plateforme". Mistral AI news, 2023-12-11. mistral.ai/...la-plateforme
  2. ^1 ^2 ^3 ^4 ^5 ^6Mistral AI. "Au Large". Mistral AI news, 2024-02-26. mistral.ai/...mistral-large
  3. ^1 ^2 ^3 ^4 ^5 ^6 ^7Mistral AI. "AI in abundance". Mistral AI news, 2024-09-17. mistral.ai/...september-24-release
  4. ^1 ^2 ^3 ^4Mistral AI. "mistralai/Mistral-Small-Instruct-2409". Hugging Face, 2024. huggingface.co/...Mistral-Small-Instruct-2409
  5. ^1 ^2 ^3Mistral AI. "params.json, mistralai/Mistral-Small-Instruct-2409". Hugging Face, 2024. huggingface.co/...params.json
  6. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Mistral AI. "Mistral Small 3". Mistral AI news, 2025-01-30. mistral.ai/...mistral-small-3
  7. ^1 ^2 ^3 ^4 ^5Mistral AI. "mistralai/Mistral-Small-24B-Instruct-2501". Hugging Face, 2025. huggingface.co/...Mistral-Small-24B-Instruct-2501
  8. ^1 ^2 ^3 ^4Mistral AI. "params.json, mistralai/Mistral-Small-24B-Instruct-2501". Hugging Face, 2025. huggingface.co/...params.json
  9. ^1 ^2 ^3 ^4 ^5 ^6 ^7Mistral AI. "Mistral Small 3.1". Mistral AI news, 2025-03-17. mistral.ai/...mistral-small-3-1
  10. ^1 ^2 ^3 ^4 ^5 ^6Mistral AI. "mistralai/Mistral-Small-3.1-24B-Instruct-2503". Hugging Face, 2025. huggingface.co/...tral-Small-3.1-24B-Instruct-2503
  11. ^1 ^2 ^3Mistral AI. "config.json, mistralai/Mistral-Small-3.1-24B-Instruct-2503". Hugging Face, 2025. huggingface.co/...config.json
  12. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Mistral AI. "mistralai/Mistral-Small-3.2-24B-Instruct-2506". Hugging Face, 2025. huggingface.co/...tral-Small-3.2-24B-Instruct-2506
  13. ^1 ^2Mistral AI. "Mistral Small 3.2". Mistral AI documentation. docs.mistral.ai/...mistral-small-3-2-25-06
  14. ^1 ^2 ^3 ^4Mistral AI. "Mistral Small Creative". Mistral AI documentation. docs.mistral.ai/...mistral-small-creative-25-12
  15. ^1 ^2 ^3 ^4 ^5 ^6 ^7Mistral AI. "Introducing Mistral Small 4". Mistral AI news, 2026-03-16. mistral.ai/...mistral-small-4
  16. ^1 ^2 ^3 ^4 ^5 ^6 ^7Mistral AI. "Models Overview". Mistral AI documentation, accessed 2026-11-18. docs.mistral.ai/models
  17. ^1 ^2Mistral AI. "mistralai/Devstral-Small-2505". Hugging Face, 2025. huggingface.co/...Devstral-Small-2505
  18. ^1 ^2Mistral AI. "mistralai/Magistral-Small-2506". Hugging Face, 2025. huggingface.co/...Magistral-Small-2506
  19. ^1 ^2Mistral AI. "mistralai/Magistral-Small-2509". Hugging Face, 2025. huggingface.co/...Magistral-Small-2509
  20. ^1 ^2Mistral AI. "mistralai/Voxtral-Small-24B-2507". Hugging Face, 2025. huggingface.co/...Voxtral-Small-24B-2507
  21. ^1 ^2 ^3Simon Willison. "Mistral Small 3". Simon Willison's Weblog, 2025-01-30. simonwillison.net/...mistral-small-3
  22. ^Maria Deutscher. "Mistral, Ai2 release new open-source LLMs". SiliconANGLE, 2025-01-30. siliconangle.com/...2-release-new-open-source-llms
  23. ^1 ^2Mistral AI. "Mistral Small 4". Mistral AI documentation. docs.mistral.ai/...mistral-small-4-0-26-03
  24. ^Mistral AI. "params.json, mistralai/Mistral-Small-3.2-24B-Instruct-2506". Hugging Face, 2025. huggingface.co/...params.json
  25. ^Alexander H. Liu et al. "Voxtral". arXiv:2507.13264, 2025. arxiv.org/...2507.13264

Improve this article

Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.

1 revision · v1 · 2,918 words · full history

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here

Reviewer note: Independent fact-check before publication on November 18, 2026: 112 claims checked against 33 cited and independent sources; 0 material and 8 minor corrections applied before publication.

Cite this page: AI Wiki. "Mistral Small." aiwiki.ai, updated 18 Nov 2026, fact-checked 18 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/mistral_small

Edit article