Mistral releases Devstral2 and Mistral Vibe CLI
European positioning and military ties
- Commenters are pleased Mistral remains European-owned and see it as strategic autonomy from US tech.
- Others argue that given existing defense contracts (e.g. EU militaries, Helsing partnership), the company is already aligned with mil-tech and will deepen that if needed.
- Some state the US has effectively “turned its back” on allies already, reinforcing the perceived need for EU AI champions.
“Vibe coding” name and philosophy
- Many dislike the “Vibe CLI” name, finding it unserious for professional work.
- Long subthread debates what “vibe coding” means:
- One camp: no reviewing code, just prompting and testing outcomes.
- Others use it more broadly for any LLM-assisted coding, even with review.
- Several note that vendors (including other major labs) are explicitly marketing “vibe coding,” which some see as encouraging sloppy, unreviewed use.
Demand for serious, review-centric tools
- Multiple users want tools that tightly integrate with IDEs, git, and diff/review workflows rather than chat-first “agents.”
- Aider is frequently cited as closest to this ideal (watch mode, auto-commits, git integration), though some still find its chat paradigm limiting.
- There’s interest in new UX paradigms: goal/milestone-based planning, better orchestration over branches, and clearer separation between AI and human edits.
Model quality, pricing, and capabilities
- Devstral 2 is seen as competitive for coding, with some placing it between mid and top-tier proprietary models.
- Early hands-on reports:
- Good at understanding codebases, finding bugs, and making localized edits.
- Strong in Python; more mixed feedback for React/JS.
- Some complaints about slow or brittle edits and occasional syntax errors.
- The announced token pricing is praised as very low; some argue pay-as-you-go now beats fixed “Pro” subscriptions. Others warn that weaker models may consume more tokens and time.
Licensing and “open source” debate
- Devstral 2’s “modified MIT” license (barring companies over €20M/month revenue) sparks long argument.
- One side: this is not “open source” or “permissive” in the standard OSI sense and misusing the term is dishonest and harmful.
- The other side: restricting only megacorps is desirable, and diluting the term is acceptable or inevitable.
- Several suggest Mistral should brand it under a custom “Mistral License” instead of “modified MIT.”
CLI, implementation, and ecosystem
- The Vibe CLI being open source (Python, Textual, Pydantic) with ACP support is welcomed; people are already packaging it (Nix, AUR) and inspecting its prompts.
- Some wish providers would contribute to existing tools (Roo, Opencode) rather than ship yet another proprietary CLI, but others argue vendors want tight optimization and ecosystem control.
- Python performance concerns are raised, though others say streaming speed issues are tool-specific, not language-limited.
Playful benchmarks and evaluation
- The familiar “SVG pelican riding a bicycle” test is used; Devstral 2 performs well, generating a coherent SVG scene.
- Long side discussion on whether such whimsical tests correlate with general capability; several claim, based on experience, that they often do, despite being originally a joke.
- Others question the value of non-realistic prompts versus practical “wine glass” style reasoning tests and worry about potential benchmark overfitting.
Local deployment and hardware
- Many are interested in running Devstral Small 2 or the full 123B model locally.
- Suggested setups range from MacBooks with large unified memory, to RTX 4090/5090, AMD 7900 XTX / AI Pro GPUs, multi-GPU 3090 rigs, and cloud rentals via llama.cpp.
- Trade-offs discussed: dense vs sparse models, VRAM requirements, power costs, and whether renting GPUs is more economical than per-token APIs.
Subscriptions, UX, and ecosystem fit
- Users miss a simple, consumer-friendly coding subscription comparable to other vendors; Mistral Code currently appears focused on enterprise/API.
- Some plan to switch from competing coding tools to Vibe for “buy European” reasons, while others remain skeptical it can match top closed models for complex work.