Key takeaways
- The only European foundation-lab coding agent — Mistral AI is backed by a €1.7B Series C at a €11.7B (~$13.8B) valuation led by ASML, with Nvidia among investors
- Launched December 9, 2025 alongside Devstral 2; as of May 2026 it defaults to Mistral Medium 3.5 (77.6% SWE-bench Verified) and can launch remote cloud agents from the CLI or Le Chat
- The tool is free and Apache 2.0; you pay per token on Mistral's API — Medium 3.5 runs $1.50/$7.50 per million tokens, Devstral 2 $0.40/$2.00
FAQ
What is Mistral Vibe?
Mistral Vibe is Mistral AI's open-source command-line coding agent that explores, modifies, and executes changes across a codebase using natural language.
How much does Mistral Vibe cost?
The CLI is free and open source (Apache 2.0). Usage is billed through Mistral's API — Mistral Medium 3.5 costs $1.50/$7.50 per million input/output tokens; Devstral 2 costs $0.40/$2.00.
What models power Mistral Vibe?
Mistral Medium 3.5 (128B dense, 256K context) is the default as of May 2026; the Devstral 2 family (123B and 24B) remains available, and the open weights can be self-hosted.
How is Mistral Vibe different from Claude Code?
Vibe is fully open source with open-weight models you can self-host, at lower per-token prices — but Claude Code is more mature, with stronger review workflows and frontier-model quality.
Executive Summary
Mistral Vibe is Mistral AI's official open-source CLI coding agent — the European entry in the foundation-lab coding agent category alongside OpenAI's Codex, Anthropic's Claude Code, Google's Gemini CLI, and Alibaba's Qwen Code. Launched December 9, 2025 with the Devstral 2 model family, it is Apache 2.0 licensed with 4.4K+ GitHub stars as of June 2026.[1][2]
The May 22, 2026 update made it more than a terminal tool: Vibe now defaults to Mistral Medium 3.5 (77.6% SWE-bench Verified) and can launch remote cloud agents — parallel, asynchronous coding tasks running in isolated sandboxes — from the CLI or from Le Chat, with local sessions "teleportable" to the cloud mid-task.[3] Mistral itself is well capitalized: a €1.7B Series C in September 2025 valued the company at €11.7B (~$13.8B), led by chip-equipment giant ASML with Nvidia among investors.[4][5]
| Attribute | Value |
|---|---|
| Company | Mistral AI |
| Founded | 2023 (Vibe: December 2025) |
| Funding | €1.7B Series C at €11.7B (~$13.8B) valuation (Sept 2025)[4] |
| GitHub Stars | 4.4K+ (June 2026)[2] |
| License | Apache 2.0[2] |
| Headquarters | Paris, France |
Product Overview
Mistral Vibe is a Python-based interactive terminal agent: you chat with your codebase in natural language, and the agent explores, edits, and runs changes across files with project-aware context scanning, multi-file orchestration, and Git integration.[6][1] The repo is 99.8% Python under Apache 2.0, with 4.4K+ stars and 526 forks; release cadence is active — v2.14.1 shipped June 8, 2026.[2]
The remote agents release (May 22, 2026) added a cloud execution layer: spawn multiple coding tasks in parallel from the CLI or Le Chat, each in an isolated sandbox, with integrations into GitHub, Linear, Jira, Sentry, Slack, and Teams for pull requests, issue tracking, and notifications. A local CLI session can be teleported to the cloud with history and approvals preserved.[3]
Key Capabilities
| Capability | Description |
|---|---|
| Conversational coding | Natural-language exploration and multi-file edits with project-aware context[2] |
| Agent profiles | default, plan, accept-edits, auto-approve modes[2] |
| Subagents + skills | Parallel task delegation and an extensible skills system[2] |
| Remote cloud agents | Asynchronous sandboxed tasks launched from CLI or Le Chat[3] |
| Multimodal input | Vision-capable image attachments via Mistral Medium 3.5[2] |
| MCP support | Model Context Protocol server configuration[2] |
Product Surfaces
| Surface | Description | Availability |
|---|---|---|
| CLI | Terminal-based interactive agent | GA[2] |
| Zed / ACP | Zed editor distribution via Agent Client Protocol | GA[2] |
| Le Chat | Launch and monitor remote coding agents from chat | GA (Pro/Team/Enterprise plans)[3] |
| Remote agents | Parallel sandboxed cloud tasks with GitHub/Jira/Slack hooks | GA[3] |
Technical Architecture
Installation:[2]
curl -LsSf https://mistral.ai/vibe/install.sh | bash
# or
uv tool install mistral-vibe
Key Technical Details
| Aspect | Detail |
|---|---|
| Deployment | Local CLI + remote cloud sandboxes[3] |
| Model(s) | Mistral Medium 3.5 (default, 128B dense, 256K context); Devstral 2 (123B), Devstral Small 2 (24B)[3][1] |
| Integrations | Zed (ACP), MCP, GitHub, Linear, Jira, Sentry, Slack, Teams[2][3] |
| Open Source | Yes — CLI Apache 2.0; Devstral Small 2 Apache 2.0, Devstral 2 modified MIT weights[2][1] |
Model benchmarks: Mistral Medium 3.5 scores 77.6% on SWE-bench Verified and 91.4 on τ³-Telecom agentic tasks, and self-hosts on as few as four GPUs; Devstral 2 scores 72.2% and Devstral Small 2 68.0%.[3][1]
Strengths
- Genuinely open stack — Apache 2.0 CLI plus open-weight models (Devstral Small 2 is Apache 2.0) you can self-host; Medium 3.5 runs on as few as four GPUs[1][3]
- Aggressive pricing — Devstral 2 at $0.40/$2.00 per million tokens; Mistral claims 7x better cost-efficiency than Claude Sonnet on real-world tasks (company claim)[1]
- Remote agents + Le Chat — Parallel sandboxed cloud tasks with session teleport, launched from terminal or chat[3]
- Enterprise tool hooks — GitHub, Linear, Jira, Sentry, Slack, and Teams integrations on remote agents[3]
- European sovereignty story — The only foundation-lab coding agent from a European lab, backed by ASML and the EU's push for AI independence[4]
- Active development — Launched December 2025; v2.14.1 by June 2026 with subagents, skills, ACP, and MCP support[2]
Cautions
- Model quality gap at the frontier — Hacker News testers reported Devstral 2 failing basic edits ("lost closing brackets or used wrong syntax") and introducing a bug during a review task; Medium 3.5's 77.6% SWE-bench Verified still trails frontier closed models[7][3]
- Small community — 4.4K+ stars versus 90K+ for Codex CLI and 105K+ for Gemini CLI; the ecosystem of extensions and community fixes is thin by comparison[2]
- Fewer review affordances — HN commenters contrasted it with Claude Code's review commands and VS Code diffs, preferring tools that disrupt existing review workflows less[7]
- The name invites skepticism — Commenters in the 349-comment launch thread objected that "vibe coding" signals unreviewed code, a liability for professional positioning[7]
- Self-hosting is expensive in practice — The flagship models are 123B–128B dense; HN commenters noted that ~120B-class models need roughly 66GB of VRAM, out of reach for most individual hardware[7]
- Value contested — One HN commenter argued Devstral 2 is comparable to DeepSeek v3.2 on its highlight task with weaker general ability at a higher price (disputed in-thread)[7]
Pricing & Licensing
| Tier | Price | Includes |
|---|---|---|
| Vibe CLI | $0 | Open-source tool (Apache 2.0)[2] |
| Mistral Medium 3.5 API | $1.50 / $7.50 per M tokens (in/out) | Default Vibe model[3] |
| Devstral 2 API | $0.40 / $2.00 per M tokens | 123B, 256K context[1] |
| Devstral Small 2 API | $0.10 / $0.30 per M tokens | 24B, Apache 2.0 weights[1] |
| Le Chat Pro / Team / Enterprise | Subscription | Remote agents from Le Chat[3] |
Licensing model: Open-source tool (Apache 2.0) + pay-per-use API or Le Chat subscription; open-weight models for self-hosting[2][1]
Hidden costs: Self-hosting the 123B–128B models requires multi-GPU hardware; remote agents require a paid Le Chat plan[3][7]
Competitive Positioning
Direct Competitors
| Competitor | Differentiation |
|---|---|
| Codex | Codex has desktop apps, 4M+ weekly developers, and frontier models; Vibe is fully open with self-hostable weights |
| Claude Code | Claude Code leads on model quality and review workflows; Vibe is cheaper per token and open source end to end |
| Gemini CLI | Gemini CLI's individual tiers sunset June 18, 2026; Vibe's open-source commitment extends to the model weights |
| Qwen Code | Both are open-source agents with open-weight models; Qwen is Alibaba's (China), Vibe is Mistral's (Europe) — sovereignty stories for opposite blocs |
| OpenCode | OpenCode is model-agnostic; Vibe is Mistral-first but pairs with first-party cloud agents |
When to Choose Mistral Vibe Over Alternatives
- Choose Mistral Vibe when: You want an open stack you can self-host, EU-based provider and data residency matter, or you want the lowest per-token prices from a first-party agent
- Choose Claude Code when: You want frontier code quality and mature review workflows
- Choose Codex when: You want the polished multi-surface OpenAI ecosystem
- Choose Qwen Code when: You want an open-weight stack with a large free tier from the Alibaba ecosystem
Ideal Customer Profile
Best fit:
- European companies with data-sovereignty or EU-residency requirements
- Teams that want to self-host an open-weight coding model behind their firewall
- Cost-sensitive teams — Devstral 2 at $0.40/$2.00 per million tokens undercuts frontier APIs[1]
- Le Chat subscribers who want to fire off background coding tasks from chat
Poor fit:
- Teams that need frontier-model code quality today — early testers hit basic-edit failures[7]
- Developers wanting a large plugin/extension ecosystem and community support
- Individuals hoping to self-host the flagship models on consumer hardware[7]
Viability Assessment
| Factor | Assessment |
|---|---|
| Financial Health | Strong — €1.7B Series C at €11.7B valuation, ASML-led with Nvidia participating[4] |
| Market Position | Niche challenger — the European option in a category led by Codex and Claude Code |
| Innovation Pace | Fast — launch to remote agents + new default model in under six months[1][3] |
| Community/Ecosystem | Early — 4.4K+ stars, 526 forks; active releases but small versus US-lab peers[2] |
| Long-term Outlook | Positive — strategic ASML backing and EU sovereignty tailwinds, contingent on closing the model-quality gap |
Mistral has the capital and strategic backing (ASML put in €1.3B for an 11% stake) to sustain Vibe indefinitely, and the EU's push for AI independence gives it a defensible customer base US labs can't easily serve.[4] The open question is model quality: the agent harness is competitive, but its models still benchmark and field-test behind the frontier.[7]
Bottom Line
Mistral Vibe is the most credible fully open foundation-lab coding agent: Apache 2.0 tool, open-weight models, self-hostable on four GPUs, and per-token prices well under frontier APIs — now with remote cloud agents launchable from the CLI or Le Chat.[1][3]
What it isn't yet is a quality leader. Early testers reported basic-edit failures, the community is two orders of magnitude smaller than Codex's or Gemini CLI's, and HN's reception mixed genuine interest with skepticism about the "vibe" branding and the models' general ability.[7][2]
Recommended for: European and sovereignty-conscious teams, self-hosters, and cost-sensitive developers who accept a quality gap for an open stack.
Not recommended for: Teams that need frontier code quality or a deep extension ecosystem today.
Outlook: Watch whether Mistral's model cadence (Devstral 2 → Medium 3.5 in five months) closes the SWE-bench gap, and whether Gemini CLI's open-source retreat sends free-tier and OSS-minded developers toward Vibe as the remaining lab-backed open agent.[3]
Research by Ry Walker Research • methodology
Sources
- [1] Introducing Devstral 2 and Mistral Vibe CLI (Mistral AI)
- [2] Mistral Vibe GitHub Repository
- [3] Vibe Remote Agents and Mistral Medium 3.5 (Mistral AI)
- [4] CNBC: Mistral Valued at $14 Billion as ASML Takes Major Stake
- [5] Mistral AI Raises €1.7B Series C (Mistral AI)
- [6] Mistral Vibe Product Page
- [7] Hacker News: Mistral Releases Devstral 2 and Mistral Vibe CLI