AI Radar

Mistral Medium 3.5

Available · Language, Multimodal, Reasoning

Mistral Medium 3.5 is a dense 128B-parameter Mistral AI model that handles instruction following, reasoning and coding in one set of weights, with text and image input and a 256k-token context window. It was released on 28 April 2026 as open weights under a Modified MIT licence and on the API, and in May 2026 became the default model in Le Chat and Mistral Vibe, in public preview. [1] [2]April 28 (2026) [3] [4] Primary source

Timeline of Mistral Medium 3.5 →

Claims and evidence

  • Released Primary source[2]April 28 (2026) [4]page header date
  • Public preview as default model in Le Chat and Mistral Vibe Primary source[3]
  • Status AvailableThe post of 2026-05-22 described it as a public preview in Le Chat and Mistral Vibe; the docs page shows it as GA. Primary source[4]GA badge [5]Generalist models [1]
  • Successor of Mistral Medium 3.1The model card says it replaces its predecessor Mistral Medium 3.1, and Magistral, in Le Chat. Primary source[1]introduction [5]Deprecated & retired models table, Alternative column
  • Change · Reasoning Adds reasoning effort set per request (none or high); the model card says it replaces Magistral in Le Chat.Compared with Mistral Medium 3.1 Primary source[1]Key Features; introduction
  • Change · Licensing Released as open weights under a Modified MIT licence; Mistral listed Medium 3.1 as Premier (commercial) with no weights link.Compared with Mistral Medium 3.1 Primary source[2]April 28 (2026) [6]
  • Change · Context length Context window grows from 128k to 256k tokens.Compared with Mistral Medium 3.1 Primary source[4]Context [6]Context
  • Change · Size Dense model with 128B parameters; Mistral did not disclose a parameter count for Medium 3.1.Compared with Mistral Medium 3.1 Primary source[1]Key Features [6]
  • Input text, image Primary source[4]Modalities [1]Key Features
  • Output text Primary source[4]Modalities [1]Key Features
  • Feature Reasoning modeConfigurable per request with the reasoning_effort parameter (none or high). Primary source[1]Key Features; Recommended Settings [2]April 28 (2026)
  • Feature Function calling Primary source[4]Features: Function Calling
  • Open weights YesModified MIT licence, with exceptions for companies with large revenue according to the model card. Primary source[1] [2]April 28 (2026)
  • Context window 256k tokens Primary source[4]Context [3]Mistral Medium 3.5 section
  • Parameters 128BDense model. Primary source[3]Mistral Medium 3.5 section [1]Key Features
  • Access consumer app, API, Open-weights download, cloud partnerLe Chat and Mistral Vibe (remote agents and Work mode on the Pro, Team and Enterprise plans), the API, Hugging Face, and NVIDIA (build.nvidia.com and NVIDIA NIM). Primary source[3]Get started

Lineage

Predecessors

Successors

No known successor.

Based on

Not derived from another model.

Variants and derived

None recorded.

Siblings

None recorded.

All ancestors

All descendants

None.

Variants

No variants recorded in this record.

Related AI Radar coverage

All model releases from Mistral AI on AI Radar →