Home Labs Mistral
Mistral
Short answer
Mistral publishes many of its models with a Hugging Face weights id, so open weights and a hosted price often appear on the same row. Large 4 is the frontier multimodal model. Devstral 2 is the coding specialist. Small 4 is the everyday cheap model. Voxtral rows cover audio.
Lab site · Filter the full catalogue to Mistral
What the top models are good for
Flagship · $0.70 in · $2.10 out · 1,048,576 tokens
Guide note · 19 Nov 2026
Mistral Large 4 is listed with a public weights id, image input, about a million tokens of context, and a blurb aimed at reasoning, coding, and agents. Pick it when open weights are a requirement and you still want a frontier-shaped model.
Open weights and a hosted API on the same row Long context Image input
Tasks: Coding , Research , Image understanding , Long documents .
What it is good for
A flagship you might later host yourself Start on the API. If the answers hold on your data, the weights id is the path to hosting. Do not assume the API price is what a GPU will cost you.
Mistral Large 4 is listed at $0.70 input and $2.10 output per million tokens, with a 1,048,576-token context window.
When a cheaper model is enough The task is a draft, a summary, or a routine coding step. Small 4 is the reviewed everyday Mistral model.
Mistral Small 4 — Mistral Small 4 is listed at $0.15 input and $0.60 output per million tokens, with a 262,144-token context window.
Outline a scanned report Image input plus a long window covers a report that arrived as pages rather than text. Ask for an outline you can check against the pages.
Mistral Large 4 is listed at $0.70 input and $2.10 output per million tokens, with a 1,048,576-token context window.
When a cheaper model is enough The scan is clean and you only need the headings and the numbers pulled out.
Mistral Small 4 — Mistral Small 4 is listed at $0.15 input and $0.60 output per million tokens, with a 262,144-token context window.
The tasks are editorial illustrations, not measured benchmarks. The prices in them are read from today's feed.
Coding · $0.40 in · $2.00 out · 262,144 tokens
Guide note · 19 Nov 2026
OpenRouter describes Devstral 2 as Mistral's open-weight coding model for agentic work, including exploring a repository. Text and file input, text output, a 256,000-token window. It is a coding tool, not a general writer.
Repository-scale coding Open weights File input
Tasks: Coding .
What it is good for
A coding agent that has to look around a repo The listing is specific: agentic coding, exploring a codebase. Point it at the failing test and the files you allow.
Devstral 2 2512 is listed at $0.40 input and $2.00 output per million tokens, with a 262,144-token context window.
When a cheaper model is enough The change is local and you want a lower rate. Qwen3 Coder Next is also open weights and is listed as a coding-agent model.
Qwen3 Coder Next — Qwen3 Coder Next is listed at $0.12 input and $0.80 output per million tokens, with a 262,144-token context window.
Explain a failing test from the files you attach File input and a coding brief. Ask for the cause and a patch, then run the test.
Devstral 2 2512 is listed at $0.40 input and $2.00 output per million tokens, with a 262,144-token context window.
When a cheaper model is enough You want a comment on a small snippet, not a repo walk.
Mistral Small 4 — Mistral Small 4 is listed at $0.15 input and $0.60 output per million tokens, with a 262,144-token context window.
The tasks are editorial illustrations, not measured benchmarks. The prices in them are read from today's feed.
Budget · $0.15 in · $0.60 out · 262,144 tokens
Guide note · 19 Nov 2026
Mistral Small 4 is the reviewed everyday Mistral model: open weights, image input, a 262,144-token window, and a low rate. The listing presents it as the small model that folds several earlier Mistral lines into one.
Drafts and routine vision Open weights Low rate
Tasks: Cheap high-volume automation , Writing , Image understanding , Coding .
What it is good for
First draft of a product email, plus the screenshot it refers to Image input means the draft can mention what is actually on screen. You still edit the email.
Mistral Small 4 is listed at $0.15 input and $0.60 output per million tokens, with a 262,144-token context window.
When a cheaper model is enough There is no screenshot, the draft can be rough, and you want the lower text rate.
DeepSeek V4 Flash 0423 — DeepSeek V4 Flash 0423 is listed at $0.09 input and $0.18 output per million tokens, with a 1,048,576-token context window.
Bulk classification that sometimes includes a picture A low rate plus image input covers a queue that is mostly text and occasionally a photo. Sample both kinds.
Mistral Small 4 is listed at $0.15 input and $0.60 output per million tokens, with a 262,144-token context window.
When a cheaper model is enough You have removed the pictures and the text is short. Nova Micro is text-only.
Nova Micro 1.0 — Nova Micro 1.0 is listed at $0.035 input and $0.14 output per million tokens, with a 128,000-token context window.
The tasks are editorial illustrations, not measured benchmarks. The prices in them are read from today's feed.
Every current row
Questions Which Mistral model should I start with? For most people, start with Mistral Large 4, then drop to the next model on this page when the task is routine. Read the cheaper-model note under each example.
Are the examples benchmarks? No. They are concrete tasks used to explain fit. The prices next to them are live figures from the feed. The judgment is a guide note.