Mistral Models Explained: Medium, Small, Large, Ministral, Codestral and More

by David Porter

Saturday, 29 August 2026 at 11:11

thumbnail_mistral-models-explained-mediu

Mistral's model catalog changes quickly. Names that dominated earlier comparisons—Mistral 7B, Mixtral, Pixtral, Magistral and Devstral—no longer describe its main cloud lineup in August 2026.

The current general-purpose range is built around Mistral Medium 3.5, Small 4, Large 3 and the three Ministral 3 sizes. Codestral remains the specialist for rapid code completion. Separate families handle embeddings, OCR, speech, moderation and safety.

This guide was checked against Mistral's model catalog on 21 August 2026. For the wider company and European strategy, read What Is Mistral AI? or visit the Mistral hub.

Current Mistral models compared

Model	Architecture	Context	Modalities	License or access
Medium 3.5	Dense, 128B	256k	Text and image input	Modified MIT weights and API
Small 4	MoE, 119B total, 6.5B active in the catalog	256k	Text and image input	Apache 2.0 weights and API
Large 3	MoE, 675B total, 41B active	256k	Text and image input	Apache 2.0 weights and API
Ministral 3 3B	Dense, 3B	256k	Text and image input	Apache 2.0 weights and API
Ministral 3 8B	Dense, 8B	256k	Text and image input	Apache 2.0 weights and API
Ministral 3 14B	Dense, 14B	256k	Text and image input	Apache 2.0 weights and API
Codestral 25.08	Specialist code model	128k	Text/code	Premier API

All six current general-purpose models support a 256k context in Mistral's hosted catalog. Local runtimes may allocate a much smaller context because the key-value cache consumes additional memory.

Mistral Medium 3.5

Medium 3.5 is Mistral's dense 128B multimodal model for agentic, reasoning and coding workloads. It was released in April 2026 as version 26.04, with the API ID mistral-medium-3-5.

It supports structured output, function calling, predicted output, document Q&A, agents, built-in tools and batch processing. Mistral also publishes its weights.

The important qualification is the license. Medium 3.5 uses a Modified MIT license rather than Apache 2.0. The license does not authorize use of the weights when the user's company—or employer—exceeded $20 million in global consolidated revenue during the preceding month. Organizations over that threshold need a commercial license or Mistral's hosted service.

Medium 3.5 costs $1.50 per million input tokens and $7.50 per million output tokens through the standard API. That makes output length a material cost variable. See the Mistral API guide for worked examples.

Mistral Small 4

Small 4 is Mistral's unified instruct, reasoning, coding and vision model. It replaces the need to select separate Small, Magistral, Pixtral or Devstral cloud models for many applications.

Its name can be misleading from a hardware perspective. Small 4 is a mixture-of-experts model with 119B total parameters. It activates only a fraction of them for each token, but all weights still require storage. Mistral describes approximately 6B active model parameters per token, or about 8B when embedding and output layers are included; its catalog rounds the active figure to 6.5B.

Small 4 supports configurable reasoning effort, text and image input, a 256k context, function calling, structured output, agents and batch inference. Its weights use Apache 2.0.

The standard API price is $0.15 per million input tokens and $0.60 per million output tokens. This combination makes Small 4 a practical first benchmark for general assistants, document work and coding agents.

Self-hosting is an enterprise-class undertaking. Mistral lists minimum configurations of 4× NVIDIA HGX H100, 2× HGX H200 or 1× DGX B200. The word "Small" describes its position in the product family, not a typical laptop deployment.

Mistral Large 3

Large 3 is a 675B-parameter mixture-of-experts model with 41B active parameters. It is multilingual, multimodal and designed for general-purpose work at high capability.

Mistral releases Large 3 under Apache 2.0, which makes it unusually permissive for a model of this scale. The practical barrier is hardware rather than access to the weights. Mistral discusses deployments on an NVL72 or an eight-GPU A100/H100 node using the appropriate checkpoint and serving stack. Ollama lists Large 3 as a cloud model rather than a normal consumer download.

The hosted price is $0.50 per million input tokens and $1.50 per million output tokens. Model pricing does not track parameter count in a simple way: Medium 3.5 is smaller in raw parameter terms but carries a higher API rate.

Large 3 is relevant when a team wants an Apache-licensed high-capacity model and has suitable AI infrastructure. It is not the logical starting point for a single workstation.

Ministral 3: 3B, 8B and 14B

Ministral 3 is the current family for edge devices, workstations and lower-cost hosted inference. All three sizes support text and vision, advertise a 256k context and use Apache 2.0.

Ministral 3 3B prioritizes low memory use and speed. Its API price is $0.10 per million input or output tokens.
Ministral 3 8B is the balanced local option and the default size in Ollama's current ministral-3 library entry. It costs $0.15 per million tokens in either direction.
Ministral 3 14B offers more capacity while remaining far easier to host than Small 4. It costs $0.20 per million input or output tokens.

Ollama's default Q4_K_M files are about 3.0GB, 6.0GB and 9.1GB respectively. These are download sizes, not complete memory requirements. Context allocation, vision processing, runtime overhead and concurrent requests require additional RAM or VRAM.

Our local Mistral guide includes the commands and hardware estimates. Related coverage is available in the hardware, GPU and Nvidia hubs.

Codestral 25.08

Codestral 25.08 is a specialist Premier model for low-latency, high-frequency coding tasks. It supports ordinary chat completions and a dedicated fill-in-the-middle endpoint.

FIM is useful when an editor knows the code before and after the cursor. Codestral receives a prefix and suffix, then generates the missing middle. Its 128k context supports substantial surrounding code, although sending an entire repository on every keystroke would be expensive and slow.

Codestral costs $0.30 per million input tokens and $0.90 per million output tokens. Current Codestral is not listed with downloadable weights.

This must be separated from two older facts:

The original Codestral 22B weights released in 2024 used the Mistral Non-Production License.
Devstral was Mistral's family for repository-scale agentic coding rather than rapid FIM autocomplete.

Devstral 2 and Devstral Small 2 are now deprecated in Mistral's cloud catalog. Local Devstral Small 2 files remain available, but new cloud integrations should evaluate Small 4 or Medium 3.5. Follow AI World Today's coding hub for broader coding-agent coverage.

OCR, audio, embeddings and safety models

Mistral's catalog extends beyond chat:

Mistral OCR 4.1 converts documents into structured content.
Voxtral models cover transcription, realtime speech recognition and text-to-speech.
Mistral Embed generates general text embeddings.
Codestral Embed generates code-focused embeddings for search and retrieval.
Shieldstral 1.0 provides an Apache-licensed safety model.
Moderation 2 provides hosted content classification.

These models use different units and endpoints. OCR can be billed per page, audio per minute or character, and embeddings per input token. Do not reuse the chat pricing table for them.

Open weights do not mean one universal license

Mistral publishes many weights, but the legal terms vary by model:

Small 4, Large 3 and Ministral 3 use Apache 2.0.
Medium 3.5 uses Modified MIT with a monthly-revenue restriction.
Current Codestral is Premier API access.
Voxtral TTS uses CC BY-NC 4.0.
Older releases can carry different terms.

Always read the specific model card and Mistral's license guidance before commercial deployment, redistribution, modification or fine-tuning. Our Mistral safety guide addresses governance and security, while the open-source AI hub covers wider questions around open-weight systems.

Which Mistral model should you choose?

A useful first selection looks like this:

Need	Starting model
Low-cost general API work	Small 4
Complex agentic or coding work	Medium 3.5
High-capacity Apache-licensed deployment	Large 3
Laptop or edge deployment	Ministral 3 3B or 8B
Stronger workstation-local model	Ministral 3 14B
IDE autocomplete and FIM	Codestral 25.08
Code retrieval embeddings	Codestral Embed
Document extraction	OCR 4.1

Benchmark with representative private tests rather than choosing from a leaderboard alone. Accuracy, latency, output length, license, regional availability and operational cost all matter.

Teams comparing ecosystems can continue with Mistral vs ChatGPT and Mistral vs DeepSeek. The ChatGPT and DeepSeek hubs provide the broader context.

What happened to Mistral 7B, Mixtral and Magistral?

Mistral 7B and Mixtral remain historically important open-weight releases, and their files can still run locally. They are not current Mistral API recommendations.

The current catalog also places Small 3.x, Medium 3.1, Large 2, Magistral, Pixtral, Devstral 2, old Ministral releases and earlier Codestral versions in deprecated or retired sections. Old tutorials may still work with downloaded weights, but API code should use current IDs.

Frequently asked questions

What is the newest general Mistral model?

Medium 3.5, released in April 2026, is newer than Small 4, which arrived in March 2026. "Newest" does not automatically mean best for every workload.

Is Mistral Small 4 a 6B model?

No. It has 119B total parameters and the catalog lists 6.5B active per token. Storage and deployment must account for the full weight set.

Are all Mistral models open source?

No. Several current models have downloadable weights, while Codestral, embeddings, OCR and other Premier models can be API-only. Licenses also differ.