Mistral AI

Mistral Medium 3.5

Provider

Mistral AI

Status

available

Context

256,000 tok

SWE-bench

77.6%

Price

$1.5 / $7.5 /MTok

Mistral Medium 3.5 is Mistral AI’s flagship working model, announced on 22 May 2026 and now the default model behind Le Chat / Vibe. Unusually for a top-tier model, it is a dense 128-billion-parameter model — not a mixture-of-experts — that unifies chat, reasoning, coding and vision in one set of weights, with reasoning effort configurable per request and self-hosting possible on as few as four GPUs (Mistral).

It is also Mistral’s move to open up the mid-tier: where the earlier Mistral Medium 3 was closed and API-only, Medium 3.5 ships as open weights under a modified MIT licence. Mistral’s headline benchmark is 77.6% on SWE-bench Verified (VentureBeat-syndicated coverage), competitive in the near-frontier tier alongside DeepSeek V4 and MiniMax M3, though below the closed leaders (Claude Opus 4.8, GPT-5.5).

Quick specs

VIEW MISTRAL MEDIUM 3.5 →

What Mistral Medium 3.5 is

Mistral Medium 3.5 is a dense 128B model — the Hugging Face repository is literally named Mistral-Medium-3.5-128B — with a 256K-token context window and a custom vision encoder trained from scratch for variable image sizes (docs.mistral.ai). Its defining design choice is consolidation: Mistral folded what had been separate reasoning (Magistral) and coding (Devstral) specialisations into a single model, with a reasoning_effort parameter that dials test-time compute up or down per request (Hugging Face).

That makes it Mistral’s agentic workhorse. It is the default model in Le Chat / Vibe, powers Vibe’s Work Mode (email, calendar and document tasks) and Code Mode, and replaced Devstral 2 as the model behind Vibe’s coding agent and CLI (Mistral). Because it is dense and relatively compact, Mistral says it runs self-hosted on as few as four GPUs — a practical option for enterprises wanting the model on their own infrastructure.

The licence

Medium 3.5’s licence is its most-discussed feature. It is open weights under a modified MIT licence — but the “modified” part is a revenue carve-out, not a ban on commercial use. Per the Hugging Face LICENSE file, the grant does not extend to companies whose global consolidated monthly revenue exceeds roughly $20 million, which must obtain a separate commercial licence from Mistral or use the hosted API. For individuals, startups, mid-market firms and most universities it behaves like ordinary MIT — commercial use, redistribution and self-hosting all allowed.

This is a meaningful change from the closed Mistral Medium 3, and it is more permissive than the Llama Community Licence, but it has drawn the same critique levelled at “open-ish” licences generally: a use restriction makes upstream licensing messier for regulated industries. It is also less permissive than its sibling Mistral Large 3, which is straight Apache 2.0.

Benchmark performance

Mistral published a narrow set of numbers at launch — strong on coding and agentic tool use, but with several common benchmarks omitted.

The picture is strong agentic and coding performance for a compact, self-hostable model, with the caveat that benchmark coverage is thin and mostly vendor-reported — a point critics have raised, noting it doesn’t top any single public leaderboard while costing more than several rivals (WinBuzzer). See best AI for coding for cross-model standings.

Pricing and access

Mistral Medium 3.5 is priced at $1.50 input / $7.50 output per million tokens on La Plateforme, with cached input around $0.15 (Artificial Analysis). That is a steep rise — roughly 3.75× on both input and output — over the Mistral Medium 3 it replaces ($0.40 / $2.00). The API model ID is mistral-medium-3-5.

The open weights are free to download from Hugging Face (subject to the modified-MIT revenue terms) and self-host on around four GPUs. It is also accessible via Le Chat / Vibe, Nvidia NIM, OpenRouter and Ollama.

How Mistral Medium 3.5 compares

• vs its sibling Mistral Large 3 — Large 3 is the bigger, more capable MoE generalist under the more permissive Apache 2.0; Medium 3.5 is smaller and dense but is the model Mistral actually ships in Vibe, tuned for agentic and coding work and cheaper to self-host.
• vs other open models — In the near-frontier open tier with DeepSeek V4, MiniMax M3 and Qwen, Medium 3.5 competes on coding and agentic reliability, but it is notably more expensive per token, and some rivals top it on raw benchmarks.
• vs the proprietary frontier — Its 77.6% SWE-bench Verified is competitive but sits below the closed leaders (Claude Opus 4.8, GPT-5.5) on the hardest coding and long-horizon agentic tasks.

Known limitations

Pricey for the tier — a sharp increase over Medium 3 and more expensive than several open rivals, without topping any single public leaderboard. Thin, mostly vendor-reported benchmarks — GPQA, MMLU-Pro, AIME, LiveCodeBench and SWE-bench Pro were not published at launch. Licence friction — the modified-MIT revenue threshold complicates use for large companies and regulated buyers compared with straight Apache 2.0. Early tooling caveat — an initial long-context bug in the Transformers config was fixed, with vLLM recommended for production (Hugging Face).

FAQ

What is Mistral Medium 3.5?

Mistral Medium 3.5 is Mistral AI’s flagship working model, announced 22 May 2026 — a dense 128B open-weight model with a 256K context window that unifies chat, reasoning, coding and vision, and powers Le Chat / Vibe.

Is Mistral Medium 3.5 open source?

It ships as open weights under a modified MIT licence. It is free for individuals, startups and most companies, but firms above roughly $20M global consolidated monthly revenue need a separate commercial licence (or must use the hosted API). The earlier Mistral Medium 3 was closed.

How much does Mistral Medium 3.5 cost?

$1.50 per million input tokens and $7.50 per million output tokens on Mistral’s API, with cached input around $0.15 — a steep increase over Mistral Medium 3’s $0.40 / $2.00. The open weights are free to self-host within the licence terms.

What model powers Mistral’s Vibe / Le Chat?

Mistral Medium 3.5 is the default model behind Le Chat / Vibe. It powers Vibe’s Work Mode and Code Mode and replaced Devstral 2 as the model behind Vibe’s coding agent and CLI.

Is Mistral Medium 3.5 a reasoning model?

Yes — it is a hybrid model with a configurable reasoning_effort parameter, so you can dial reasoning compute up for hard problems or off for fast, cheap responses, all from one set of weights.

How does Mistral Medium 3.5 compare to Mistral Large 3?

Mistral Large 3 is the larger, more capable mixture-of-experts generalist under Apache 2.0. Medium 3.5 is a smaller dense model, more expensive per token but tuned for agentic and coding work, easier to self-host (~4 GPUs), and it is the model Mistral actually ships in Vibe.

Last verified 19 June 2026. The 22 May 2026 announcement, dense-128B architecture, modified-MIT licence, pricing and the 77.6% SWE-bench Verified figure are from Mistral’s own materials, corroborated by Hugging Face, Artificial Analysis and OpenRouter. Mistral’s docs version-stamp the model April 2026 (model-card slug 26-04); benchmark coverage is thin and largely vendor-reported, so figures should be confirmed against current leaderboards before relying on them.