GPT-56T 861 —
MUSE-SPK 837 +0.2%
GPT-56SC 790 -4.6%
GLM-5 781 -0.4%
CL-OP55X 780 -5.1%
GROK-46H 780 -5.1%
QWEN-38X 748 -9.2%
GPT-6A 743 -9.4%
KIMI-K3X 742 —
CL-FAB5H 698 -6.1%
CL-OP5H 675 -6.2%
GEM-38FH 672 -0.7%
CL-OP5X 670 -5.5%
CL-OP55H 668 —
CL-OP46H 657 -5.9%
CL-OP47H 648 -6.1%
GPT-56S 618 -0.6%
GEM-37FH 610 -7.2%
GEM-36FH 593 —
CL-OP48H 588 —
CL-OP47 581 -0.2%
GEM-35FH 580 —
GPT-55H 541 -7%
INKL 531 —
GEM-31P 512 -0.2%
CL-OP46 498 +0.4%
GEM-3P 498 -0.2%
CL-OP48 492 +0.4%
GPT-52 464 —
GPT-55 423 —
GPT-56T 861 —
MUSE-SPK 837 +0.2%
GPT-56SC 790 -4.6%
GLM-5 781 -0.4%
CL-OP55X 780 -5.1%
GROK-46H 780 -5.1%
QWEN-38X 748 -9.2%
GPT-6A 743 -9.4%
KIMI-K3X 742 —
CL-FAB5H 698 -6.1%
CL-OP5H 675 -6.2%
GEM-38FH 672 -0.7%
CL-OP5X 670 -5.5%
CL-OP55H 668 —
CL-OP46H 657 -5.9%
CL-OP47H 648 -6.1%
GPT-56S 618 -0.6%
GEM-37FH 610 -7.2%
GEM-36FH 593 —
CL-OP48H 588 —
CL-OP47 581 -0.2%
GEM-35FH 580 —
GPT-55H 541 -7%
INKL 531 —
GEM-31P 512 -0.2%
CL-OP46 498 +0.4%
GEM-3P 498 -0.2%
CL-OP48 492 +0.4%
GPT-52 464 —
GPT-55 423 —
← Back to feed

Mistral Medium 3.5: Open-Weights 128B Posts 77.6% SWE-Bench and Ships Cloud Coding Agents

Mistral today released Mistral Medium 3.5 in public preview, its first flagship merged model and the new default for both Le Chat and the Vibe coding agent CLI. The model ships as open weights under a modified MIT license, self-hostable on four GPUs.

The Model

Medium 3.5 is a 128B dense model with a 256k context window. Unlike Mistral’s previous specialist models — Devstral 2 for code, Pixtral for vision — Medium 3.5 merges instruction-following, reasoning, and coding into a single set of weights. Reasoning effort is configurable per request: the same model handles a quick chat reply or a long agentic run.

Mistral trained the vision encoder from scratch to handle variable image sizes and aspect ratios, extending the model into multimodal territory without a separate deployment.

Benchmark Numbers

BenchmarkScore
SWE-Bench Verified77.6%
tau³-Telecom91.4%

77.6% on SWE-Bench Verified beats Devstral 2 and Qwen3.5 397B. For context: Claude Opus 4.6 sits at 80.8%, GPT-5.4 at 77.2%, and Qwen3.6-27B at 77.2%. Medium 3.5 ties GPT-5.4 on SWE-Bench while running on a fraction of the infrastructure. Tau³-Telecom at 91.4% is strong in the agent reliability category — above the 91.9% reported for Claude Opus 4.6 on tau2-Telecom and within the range of purpose-built agentic models.

Vibe Remote Agents

The bigger operational shift is cloud-hosted coding agents. Vibe sessions now run remotely, multiple at once, notifying the user when work completes. Previously, Vibe ran locally.

Capabilities at launch:

  • Start sessions from the Vibe CLI or from Le Chat
  • Live inspection: file diffs, tool calls, progress states, approval prompts surfaced in real time
  • Local CLI sessions can be teleported to the cloud mid-session, carrying history, state, and pending approvals
  • GitHub integration for pull request creation, Linear and Jira for issue tracking, Sentry for incident response

A new Work mode in Le Chat runs on Medium 3.5 for multi-step research, analysis, and cross-tool tasks — the consumer-facing equivalent of the Vibe agent runtime.

Context

Mistral raised $830M in March 2026 at a $14B valuation targeting 13,800 Nvidia GB300 GPUs. Medium 3.5 is the first major model to come out of that compute expansion. API pricing for Medium 3.5 was not announced at launch. Medium 3 API pricing is $0.40/M input, $2.00/M output.

Open weights at this benchmark tier are materially significant. The previous generation of models at 77% SWE-Bench was firmly proprietary. Medium 3.5 gives enterprise teams a self-hosted path to GPT-5.4-equivalent coding performance without vendor lock-in.