What is Mistral 3?
Mistral 3 is a family of four open-weight models. Ministral 3 comes in 3B, 8B and 14B sizes, each in base, instruct and reasoning versions, built to run on PCs, laptops and edge devices. Mistral Large 3 is a 675B mixture-of-experts model with 41B parameters active per token, for chat, documents and tool use. All four read text and images, write text and handle 256K tokens.
They are still on sale, but no longer Mistral's newest: Mistral Medium 3.5 is now its top model and the default in Le Chat, and Mistral Small 4 fills the gap between Ministral 3 14B and Large 3.
Is Mistral 3 still current?
Yes. All four Mistral 3 models are still sold on Mistral's API, but Mistral Medium 3.5 has since become Mistral's top model.
- Status
- Ministral 3 3B, 8B, 14B and Mistral Large 3 are generally available, and none is on Mistral's deprecation list.
- Newer flagship
- Mistral Medium 3.5, a 128B dense model at $1.5 / $7.5 per million tokens, is now Mistral's top model and the default in Le Chat and Vibe.
- Middle of the range
- Mistral Small 4, a 119B mixture of experts with built-in reasoning, sits between Ministral 3 14B and Large 3.
- What it replaced
- Mistral points users of the retired Ministral 3B and 8B, Pixtral 12B and Mistral Large 2.0 to the Mistral 3 models.
Who is Mistral 3 best for?
Mistral 3 is best for developers who want open-weight models they can self-host under Apache 2.0, from Ministral 3 on a single GPU to Mistral Large 3 on an eight-GPU server, or cheap API calls.
Self-hosting under Apache 2.0
Every size can be downloaded, fine-tuned and shipped inside a product with no custom licence terms and no per-token fees.
Local apps on one GPU
Ministral 3 3B fits in 8GB of VRAM in FP8, the 8B in 12GB and the 14B in 24GB, with less after further quantization.
Cheap high-volume API jobs
Ministral 3 3B costs $0.1 per million tokens on Mistral's API, and batch jobs cost half that.
Multilingual assistants
Mistral says the family handles more than 40 languages, and it singles out Large 3 for conversations outside English and Chinese.
What's included with Mistral 3?
Image input on every size, reasoning variants of Ministral 3, tool calls and JSON output, half-price batch jobs, compressed checkpoints for self-hosting, and availability on the main clouds.
- Image input on every size
- All four models read images as well as text; Large 3 uses a 2.5B vision encoder and each Ministral model a 0.4B one.
- Reasoning variants
- Each Ministral 3 size also comes as a reasoning model that thinks longer before it answers; Large 3 has base and instruct versions only.
- Tool calls and JSON output
- Native function calling and structured outputs on every model, plus Mistral's Agents API and built-in tools on Large 3.
- Batch and Priority tiers
- Batch jobs cost half the standard price, and the Priority tier costs more for committed service levels.
- Compressed checkpoints
- An NVFP4 build of Large 3 runs on one node of eight A100 or H100 GPUs, and Mistral publishes GGUF files of the Ministral models.
- Runs on the main clouds
- Sold on Mistral's API and through Amazon Bedrock, Azure Foundry, IBM watsonx, OpenRouter, Fireworks and Together AI.
How much does Mistral 3 cost?
Mistral 3's weights are free to download under Apache 2.0; on Mistral's API, Ministral 3 costs $0.1 to $0.2 per million tokens and Mistral Large 3 costs $0.5 per million input and $1.5 per million output tokens.
- Smallest model, for edge and local use
- Batch: $0.05 / $0.05
- Priority: $0.18 / $0.18
- Base, instruct and reasoning versions with image input
- Batch: $0.075 / $0.075
- Priority: $0.263 / $0.263
- Largest Ministral 3 model
- Batch: $0.1 / $0.1
- Priority: $0.35 / $0.35
- 675B-parameter mixture of experts, 41B active
- Batch: $0.25 / $0.75
- Priority: $0.88 / $2.63
These are Mistral's own API prices in US dollars. Cached input tokens cost 90% less, and EU or global regional endpoints add 10% on supported models. All four models are free to download from Hugging Face under the Apache 2.0 license, and Mistral's free plan includes $10 a month in API credits. The models are also sold through Amazon Bedrock, Azure Foundry, IBM watsonx, OpenRouter, Fireworks and Together AI at those providers' prices. Full pricing on mistral.ai. We checked these prices on .
How do readers rate Mistral 3?
No one has rated Mistral 3 yet. Be the first.
Sign in to rate. One rating per account.
Mistral 3 pros and cons
Apache 2.0 weights and low prices are the draw, but Large 3 trails Qwen and DeepSeek, needs an eight-GPU server, and there is no model between 14B and 675B.
Pros
- All four models ship under Apache 2.0, which users welcome as simpler and more open than the custom licences on some other open models.
- Ministral 3 14B holds up well on local hardware: early users rated its output close to or above Gemma 3 27B and the older 24B Mistral Small, on lighter hardware.
- Low API prices, from $0.1 per million tokens for Ministral 3 3B to $0.5 / $1.5 for Large 3, which users call good enough for the money on routine jobs.
- Open weights let companies keep data on their own servers, and several European users say that is why they pick Mistral over US-hosted APIs.
Cons
- It trails the best open rivals: users put Large 3 roughly level with the older DeepSeek V3.1 and behind Qwen3-235B, and found Ministral 3 3B weaker than Qwen's 4B at following instructions.
- Nothing sits between 14B and 675B, and Large 3 needs a full eight-GPU server even in its compressed NVFP4 form, which rules it out for home and small-office setups.
- Several users saw the Ministral reasoning variants loop, overthink or break tool calls in local runners; Mistral recommends its own system prompt and a temperature of 1 for them.
- Mistral says Large 3 is not a dedicated reasoning model and can lag behind vision-first models on image tasks.
Is Mistral 3 open source?
Yes. Mistral 3's code is public under the Apache 2.0 licence, so you can run it on your own server.
What are the best Mistral 3 alternatives?
Mistral 3 is an open source alternative to LFM 2.5. Gemma 4 and Qwen3.5 Small do the same job and are open source too.
LFM 2.5FreeLiquid AI's open-weight model family built for edge devices, with tool calling.
Gemma 4Open SourceChoose it when you need small open models that also take audio (the E2B, E4B and 12B sizes) or want free API calls, since Google charges nothing for Gemma 4 on the Gemini API's free tier.
Qwen3.5 SmallOpen SourceQwen3.5's 0.8B to 9B open multimodal models.
Mistral Small 4Open SourceChoose it when you want one newer Mistral model with switchable reasoning, image input and a 256K context, sized between Ministral 3 14B and Large 3; it costs $0.15 / $0.6 per million tokens.
DeepSeek V4.1 FlashOpen SourceChoose it when you need a 1M-token context and lower hosted prices than Large 3: $0.30 / $1.20 per million tokens at peak and half that off-peak, with MIT-licensed weights.
Questions about Mistral 3
What is Mistral 3?
Mistral 3 is Mistral AI's open-weight model family: three Ministral 3 models (3B, 8B, 14B) and the 675B Mistral Large 3, all under Apache 2.0, for developers who self-host or want cheap API calls.
Who is Mistral 3 for?
Mistral 3 is best for developers who want open-weight models they can self-host under Apache 2.0, from Ministral 3 on a single GPU to Mistral Large 3 on an eight-GPU server, or cheap API calls.
Is Mistral 3 free?
Yes, the Mistral 3 weights are free to download and use under Apache 2.0. Running them on Mistral's API is paid per token, from $0.1 per million tokens for Ministral 3 3B to $0.5 input and $1.5 output for Mistral Large 3; the free plan includes $10 a month in API credits.
How much does Mistral 3 cost?
Mistral 3's paid plans are Ministral 3 (3B) at $0.1 / $0.1 per million input / output tokens, Ministral 3 (8B) at $0.15 / $0.15 per million input / output tokens, Ministral 3 (14B) at $0.2 / $0.2 per million input / output tokens and Mistral Large 3 at $0.5 / $1.5 per million input / output tokens.
Is Mistral 3 open source?
Yes. Mistral 3's code is public under the Apache 2.0 licence, so you can run it on your own server.
What are the best Mistral 3 alternatives?
The closest alternatives on Recaply are LFM 2.5, Gemma 4, Qwen3.5 Small, Mistral Small 4 and DeepSeek V4.1 Flash.
Where can I try Mistral 3?
On its official website, mistral.ai.
Is Mistral 3 the same as Mistral Small 3?
No. Mistral 3 is the family of Ministral 3 and Mistral Large 3. Mistral Small 3, 3.1 and 3.2 were an earlier 24B model line; Mistral has deprecated Small 3.2 and points new projects to Mistral Small 4.
Which Ministral 3 model fits on my GPU?
Mistral says Ministral 3 3B fits in 8GB of VRAM in FP8, the 8B in 12GB and the 14B in 24GB, and less after further quantization. All three read images and handle a 256K-token context.
Can I run Mistral Large 3 on my own hardware?
Yes, on a server. Its NVFP4 checkpoint runs on a single node of eight A100 or H100 GPUs with vLLM, and the FP8 version on a single node of B200s or H200s. It has 675B parameters, 41B of them active per token.
Does Mistral Large 3 have a reasoning mode?
No. Mistral released Large 3 in base and instruct versions and says it is not a dedicated reasoning model. For step-by-step reasoning, use a Ministral 3 reasoning variant, or Mistral Small 4, which takes a reasoning_effort setting on the API.
Our verdict
Mistral 3 is a sound pick when an Apache 2.0 licence and self-hosting matter more than top scores.
The Ministral 3 models run on one ordinary GPU and cost $0.1 to $0.2 per million tokens on the API, but users find Large 3 behind Qwen and DeepSeek, and Mistral's own Small 4 and Medium 3.5 now cover the middle and the top.
- Skip it for
- Hard coding or agent work where you need the strongest open model, or a mid-size model to run at home.
- Check first
- Test the Ministral reasoning variants with Mistral's own system prompt before you judge them.
Mistral Medium 3.5
Step 3.7 Flash
1-bit Bonsai
Perceptron Mk1
Mistral AI
SciSpace