SHELFCRITTER AI Digest

MODEL

Mistral Medium 3.5

modeltopic-notemistral

Overview

Mistral Medium 3.5 is a 128B dense multimodal model released by Mistral on May 1, 2026, as part of a dual launch alongside Vibe remote agents. The model is Mistral’s first substantive 2026 release following the March–April Large 3 cycle and represents a push into multimodal reasoning with a focus on dense architectures rather than mixture-of-experts patterns.

Timeline

  • 2026-05-03-AI-Digest — Mistral ships Medium 3.5 (128B dense multimodal) alongside Vibe remote agents. Mistral’s published headline: 77.6% SWE-Bench Verified. Not independently corroborated on public leaderboard at writing; treat as Mistral’s own benchmark unless and until it shows up on the maintained leaderboard.

Key Specs

  • Architecture: 128B dense (not mixture-of-experts)
  • Capabilities: multimodal (vision + text)
  • SWE-Bench Verified: 77.6% (claimed by Mistral, unverified)

Key Developments

  1. Dense Multimodal Focus: Medium 3.5’s 128B dense architecture continues Mistral’s preference for dense models over the mixture-of-experts pattern dominating the broader 2026 frontier cohort (DeepSeek V4, GPT-5.4, Gemini 3.1).

  2. Unverified Benchmark Claim: Mistral’s 77.6% SWE-Bench Verified headline did not appear on the public leaderboard at the time of this digest’s writing. The claim is credible given Mistral’s earlier Large 3 performance, but the public verification gap is worth noting for procurement purposes.

  3. Companion Vibe Launch: Medium 3.5 ships as part of the same announcement as Vibe, Mistral’s new remote agents product, suggesting the model is optimized for agentic use cases rather than broad general-purpose deployment.

  4. API Pricing: Available through Mistral’s API at $1.50 per 1M input tokens and $7.50 per 1M output tokens on Pro/Team/Enterprise plans.

See also: Mistral, Vibe, MOC - Developer Tools