Developer
Mistral AI

Mistral Medium 3.5

Mistral AI's multimodal model for agentic and coding work: a dense 128B model with a 256K context window, released April 2026 under a modified MIT license.

Updated

Released
Apr 28, 2026
Type
Multimodal Language Model
Context window
256K tokens
Pricing
$1.50 / $7.50 per Mtok
License
Modified MIT (open weights)
🌐 Official Website📚 Documentation💰 Pricing & API
On this page

Overview

Mistral Medium 3.5 is Mistral AI's frontier-class multimodal model, "optimized for agentic and coding use cases." Mistral's changelog dates the release to April 28, 2026 (the model card carries version 26.04); the public announcement followed on May 22, 2026, presenting the model in public preview alongside Vibe remote agents.

It is a dense 128B-parameter model with a 256K context window, and Mistral calls it "our first flagship merged model" — instruction-following, reasoning, and coding consolidated into a single set of weights, with reasoning effort configurable per request rather than split across separate model variants.

The most consequential change from the previous generation is licensing. Mistral Medium 3.1 was a proprietary Premier-tier model. Medium 3.5 ships as open weights under a modified MIT license.

Medium 3.5 is not Mistral's overall flagship — that is Mistral Large 3. It is the tier Mistral tunes hardest for agentic and software-engineering work.

Capabilities

  • Merged reasoning and instruction weights: One model handles quick chat and long agentic runs. Mistral notes that "reasoning effort is now configurable per request."
  • Agentic coding: The model behind Mistral's Vibe remote agents, which run asynchronous, parallel coding sessions in isolated cloud sandboxes.
  • Multimodal input: A custom-trained vision encoder handles variable image sizes and aspect ratios rather than forcing a fixed crop.
  • 256K context: Twice the 128K window of Mistral Medium 3.1.
  • Open weights: Downloadable and self-hostable under a modified MIT license.
  • Full API surface: Chat completions, function calling, agents and conversations, built-in tools, structured outputs, predicted outputs, and prefix completion.

Technical Specifications

  • Model ID: mistral-medium-3-5-26-04 (rolling alias: mistral-medium-latest)
  • Version: 26.04
  • Architecture: Dense transformer, 128B parameters
  • Context window: 256K tokens
  • Modalities: Text and image input; text output
  • Vision encoder: Custom-trained, accepts variable image sizes and aspect ratios
  • Reasoning: Configurable effort per request
  • License: Modified MIT (open weights)
  • Availability: Mistral API (La Plateforme), Le Chat, Mistral Vibe, and self-hosting from published weights

Mistral has not published a knowledge cutoff date or a maximum output token count for Medium 3.5.

Where Medium 3.5 sits in Mistral's lineup

Mistral Medium 3.5Mistral Large 3
Model IDmistral-medium-3-5-26-04mistral-large-2512
ReleasedApril 28, 2026December 2, 2025
ArchitectureDense, 128BSparse MoE, 675B total / 41B active
Context window256K256K
MultimodalYesYes
Input price$1.50 / MTok$0.50 / MTok
Output price$7.50 / MTok$1.50 / MTok
LicenseModified MITApache 2.0 per the announcement; "Open v25.12" per the docs model card

Two things stand out. First, Mistral Large 3 — not Medium 3.5 — is the company's general-purpose flagship: a 675B total / 41B active sparse Mixture-of-Experts model trained from scratch on 3,000 NVIDIA H200 GPUs, and Mistral's first MoE since the Mixtral series.

Second, and counterintuitively, Medium 3.5 costs three to five times more per token than the flagship. A plausible architectural explanation: Large 3 activates 41B parameters per token, while dense Medium 3.5 activates all 128B on every token.

Note also that the license label for Large 3 is inconsistent across Mistral's own materials — the launch announcement says Apache 2.0, while the documentation model card records "Open v25.12."

Use Cases

  • Autonomous coding agents: The model powering Vibe remote agents — cloud coding sessions that run asynchronously and in parallel, integrating with GitHub, Linear, Jira, Sentry, Slack, and Teams.
  • Long-context software work: 256K tokens is enough for large repositories, dependency trees, and multi-file refactors in a single prompt.
  • Tool-using enterprise agents: Function calling, built-in tools, and structured outputs across multi-step business workflows.
  • Document and chart understanding: Multimodal input for scanned documents, screenshots, diagrams, and figures.
  • Sovereign and on-premises deployment: Open weights under a modified MIT license make EU-jurisdiction and air-gapped deployments possible without a proprietary agreement.
  • Cost-tiered routing: Configurable reasoning effort lets a single deployment serve both cheap chat turns and expensive agentic runs.

Performance / Benchmarks

Mistral has published two headline figures for Medium 3.5. Both are first-party, reported by Mistral:

  • SWE-Bench Verified: 77.6% — "ahead of Devstral 2 and models like Qwen3.5 397B A17B."
  • τ³-Telecom: 91.4 — Mistral's cited measure of agentic capability.

Mistral has not published a broader benchmark table for Medium 3.5 in its model card or announcement, so no MMLU, GPQA, or math scores are listed here.

For context on the flagship: Mistral states that Mistral Large 3 debuted at #2 in the open-source non-reasoning category (#6 among open-source models overall) on the LMArena leaderboard. LMArena is a third-party human-preference leaderboard, not a Mistral evaluation.

Limitations

  • Not the flagship: Mistral Large 3 is the general-purpose flagship. Medium 3.5 is tuned for agentic and coding work, and Mistral has not published head-to-head general-knowledge comparisons between the two.
  • Priced above the flagship: At $1.50/$7.50 versus Large 3's $0.50/$1.50, routing general work to Medium 3.5 costs more than routing it to a larger model.
  • Dense, so no MoE sparsity discount: All 128B parameters activate on every token, which sets a floor on both serving cost and self-hosted memory.
  • Thin published benchmark record: Two reported numbers, both first-party, with no independent evaluation cited by Mistral.
  • No published knowledge cutoff: Mistral does not state a training data cutoff for Medium 3.5.
  • Modified, not standard, MIT: The weights are open, but the license is a modified MIT rather than plain MIT or Apache 2.0. Read the terms before assuming permissive reuse.
  • Public preview framing: Mistral's announcement described the model as "available in public preview," so behavior and availability may still shift.

Pricing & Access

API pricing

  • Input: $1.50 / MTok
  • Output: $7.50 / MTok

By comparison, Mistral Large 3 is $0.50 / MTok input and $1.50 / MTok output.

Access paths

  • Mistral API (La Plateforme) — mistral-medium-3-5-26-04, or the mistral-medium-latest alias
  • Le Chat — including Work mode for complex tasks
  • Mistral Vibe — CLI and cloud remote agents, on Pro, Team, and Enterprise plans
  • Self-hosted — open weights on Hugging Face under a modified MIT license

Migrating from Medium 3.1

mistral-medium-2508 was marked deprecated on May 22, 2026, with retirement scheduled for August 31, 2026 and Mistral Medium 3.5 named as its replacement. The move brings a 128K → 256K context increase and open weights, but pricing and behavior both differ — re-test prompts rather than assuming a drop-in swap.

Ecosystem & Tools

  • Mistral AI Documentation: API reference, model cards, and the deprecation schedule.
  • Mistral Models Overview: Live catalog with versions, licenses, and retirement dates.
  • Hugging Face: Open weights for Medium 3.5, Large 3, Small 4, Ministral 3, Devstral 2, and Voxtral.
  • Mistral Vibe: CLI and cloud remote agents for asynchronous coding sessions, with GitHub, Linear, Jira, Sentry, Slack, and Teams integrations.
  • Le Chat: Mistral's assistant interface, including Work mode.

Community & Resources

Frequently Asked Questions

Mistral AI's frontier-class multimodal model, optimized for agentic and coding use cases. It is a dense 128B-parameter model with a 256K token context window, and Mistral describes it as its "first flagship merged model" — instruction-following, reasoning, and coding folded into a single set of weights.
Mistral's changelog dates the release to April 28, 2026, and the model card carries version 26.04. Mistral's public announcement, on May 22, 2026, described the model as available in public preview alongside Vibe remote agents.
Mistral Medium 3.1 (mistral-medium-2508) was released on August 12, 2025 with a 128K context window. It was marked deprecated on May 22, 2026, with retirement scheduled for August 31, 2026, and Mistral Medium 3.5 named as its replacement.
Yes. It is released as open weights under a modified MIT license. This is a change from Mistral Medium 3.1, which was a proprietary Premier-tier model with no published weights.
The dated model ID is mistral-medium-3-5-26-04. The rolling alias mistral-medium-latest also points at the current Medium tier.
$1.50 per million input tokens and $7.50 per million output tokens via the Mistral API.
No. Mistral Large 3 (mistral-large-2512, released December 2, 2025) is the overall flagship: a 675B total / 41B active sparse Mixture-of-Experts model with a 256K context window. Medium 3.5 is the mid-tier model tuned specifically for agentic and coding work.
It does — $1.50/$7.50 for Medium 3.5 against $0.50/$1.50 for Large 3. Mistral has not explained the inversion publicly. One architectural reading: Large 3 is a sparse MoE that activates only 41B of its 675B parameters per token, while Medium 3.5 is a 128B dense model that activates every parameter on every token, so Medium 3.5's serving cost per token can be higher despite the smaller nameplate size.
Mistral reports 77.6% on SWE-Bench Verified, ahead of Devstral 2 and of models like Qwen3.5 397B A17B, and 91.4 on τ³-Telecom for agentic capability. These are Mistral's own reported figures.
Yes. Reasoning effort is configurable per request, so the same weights serve a quick chat reply or a long agentic run.
Yes. It handles multimodal input using a custom-trained vision encoder that accepts variable image sizes and aspect ratios.

Related Models

Command A+

Cohere's first Mixture-of-Experts model, released May 2026. 218B total / 25B active parameters, vision input, reasoning, and agentic tool use under Apache 2.0.

Llama 4

Meta's last open-weight frontier release: Scout and Maverick, natively multimodal Mixture-of-Experts models with 17B active and up to a 10M token context.

GPT-5.6

OpenAI's flagship family — Sol, Terra and Luna — generally available since July 9, 2026. All three share a 1,050,000 token context and a February 2026 cutoff.

Claude Sonnet 5

Anthropic's best mix of speed and intelligence — the most agentic Sonnet yet, reaching near-Opus quality on coding and agentic work at Sonnet cost.

DeepSeek V4

DeepSeek's open-weight MoE flagship: V4-Pro (1.6T total / 49B active) and V4-Flash (284B / 13B), with hybrid sparse attention, a 1M context, and an MIT license.

Gemma 4

Google's open-weight Gemma 4 family — five Apache 2.0 models from 2.3B to a 30.7B dense flagship, spanning 140+ languages and native audio, vision, and text.

Explore More Models

Discover other AI models and compare their capabilities.

Browse Models← Back to Models