Mistral Medium 3.5: Flagship 128B Open-Weights Model Powers Remote Vibe Coding Agents
Summary
- • Mistral launches 128B Medium 3.5 with open weights (modified MIT), 77.6% SWE-Bench score, self-hostable on 4 GPUs
- • Vibe Remote Agents bring cloud-based async coding sessions with GitHub, Linear, Jira, and Sentry integrations
- • Le Chat gains Work Mode for complex multi-step agentic tasks powered by the new model
- • Medium 3.5 replaces Devstral 2 as default in both Le Chat and Mistral Vibe CLI
Details
Dense 128B model with 256k context, unified instruction/reasoning/coding weights, configurable reasoning effort per request
Not a MoE architecture — a single dense weight set handles quick chat and deep agentic runs via a runtime reasoning effort parameter, avoiding the need to maintain or switch between model variants.
77.6% on SWE-Bench Verified (ahead of Devstral 2 and Qwen3.5 397B A17B); 91.4 on τ³-Telecom
SWE-Bench Verified measures real GitHub issue resolution — the most grounded public coding benchmark. Outperforming a 397B MoE model at 128B dense parameters is a meaningful efficiency-per-parameter claim.
Vision encoder trained from scratch supports variable image sizes and aspect ratios
Most multimodal models crop or resize inputs to fixed dimensions, degrading quality. Variable-size encoding preserves spatial information for diagrams, UI screenshots, and technical schematics.
Vibe Remote Agents: cloud async coding sessions in isolated sandboxes, parallel execution, local-to-cloud session teleportation
Sessions preserve history, task state, and approvals across local and cloud execution. Agents open GitHub PRs and notify users on completion, so engineers review results rather than every step.
Integrates with GitHub, Linear, Jira, Sentry, Slack, and Teams for end-to-end async engineering automation
Targets routine high-volume tasks: module refactors, test generation, dependency upgrades, CI investigations, bug fixes — work that takes developer time without requiring developer judgment.
Medium 3.5 replaces Devstral 2 as default across Le Chat and Vibe CLI; open weights under modified MIT license
Signals Mistral consolidating toward a single general-purpose flagship rather than a tiered stack of specialized models. MIT-style license enables commercial self-hosting on 4 GPUs.
New Tech = novel capability/architecture, Research = benchmark results, Product Launch = new product/feature, Infrastructure = integration detail, Industry Update = business/platform change
What This Means
Mistral is making a direct push into the agentic developer tools market, competing with GitHub Copilot Workspace and similar async coding agents by combining a competitive open-weights model with cloud execution infrastructure. The open-weights release under a modified MIT license gives self-hosting teams a credible alternative to closed APIs at manageable hardware cost (4 GPUs). For engineering teams, the Vibe remote agent pipeline — spanning task intake through PR creation and Slack notification — represents a complete async automation loop for routine engineering tasks.
Sentiment
Broadly excited about benchmarks, open-weights accessibility, and agent integration, with pockets of skepticism on performance vs. smaller rivals
“Mistral Medium 3.5 128B is HERE and it just made their own coding model obsolete — beats Devstral 2 on SWE-Bench Verified, unified instruct/reasoning/coding weights, hands-on coding test”
“Mistral-medium-3.5 is live on Merge Gateway... 128B open-weights model... 77.6% on SWE-bench Verified. One of the only frontier models you can self-host on 4 GPUs”
“Mistral Medium 3.5 vient de sortir : 128B dense, open weights, 77,6 % sur SWE-Bench”
“I'm waiting for @OpenRouter to add... Mistral Medium 3.5 128B dense... pit it against Qwen3.6 27B... no-contest... old pretrained backbone... Mistral has some serious catching up to do”
Split
~80/20 excited/skeptical — most praise self-hosting and coding benchmarks, doubters highlight lag vs. efficient small models like Qwen.