Mistral AI’s latest drop, Medium 3.5, landed on April 29—and instead of fanfare it mostly met a chorus of “meh.” The Paris-based lab shipped a dense, 128‑billion-parameter model with agentic features, but critics say the release doesn’t match its ambitions—except for one point that keeps the company relevant.
What Mistral actually released
- Medium 3.5: a 128B-parameter dense model intended as a new flagship.
- Mistral Vibe CLI: cloud-based “remote coding agents” that can run parallel coding sessions, push pull requests to GitHub, and operate without an active terminal session.
- Work Mode in Le Chat: an expanded ChatGPT‑style interface that can execute multi‑step autonomous workflows—email triage, research synthesis, and cross‑tool jobs.
A tidy engineering win — and some weak benchmark numbers
Mistral folded three prior models (Medium 3.1, Magistral, and Devstral 2) into a single set of weights, with configurable “reasoning effort” per request—a meaningful backend simplification. But on public benchmarks the new model underwhelms: Medium 3.5 scores 77.6% on SWE‑Bench Verified (which tests whether a model can generate working patches for real GitHub issues) and 91.4% on τ³‑Telecom (agentic tool use in specialized settings). Third‑party leaderboards haven’t fully ranked the model yet.
Price and competition
Mistral’s pricing is notable: $1.50 per million input tokens and $7.50 per million output tokens—placing it near the cost of some closed models that outperform it on many benchmarks. Meanwhile, Alibaba’s Qwen 3.6 (27B parameters) scores 72.4% on SWE‑Bench Verified, ships under Apache‑2.0 (downloadable and self‑hostable), and occupies top spots on open‑source leaderboards alongside China’s GLM (Zhipu AI) and Xiaomi’s MiMo‑V2. Those projects currently dominate open‑source rankings, making Mistral’s offering comparatively expensive and not clearly superior in performance.
Community reaction
Responses ranged from dismissive to cautiously supportive. Critics called out Mistral for high costs and mediocre leaderboard performance; machine‑learning professor Pedro Domingos questioned whether Europe should be represented by what he called “a laughingstock.” Others, like developer Michal Langmajer, expressed relief that a non‑US, non‑Chinese lab is still pushing frontier LLM work, while urging European teams to “level up.”
The counterargument: strategic positioning over leaderboard wins
Supporters and some enterprise customers argue this is a durability play, not just a race for benchmark supremacy. Open weights let organizations download, fine‑tune, and self‑host models—a critical feature for institutions that must keep data on‑premises. That pitch has real traction: as Decrypt reported last December, HSBC signed a multi‑year deal to self‑host Mistral models. For many European banks, governments, and GDPR‑sensitive enterprises, a Paris‑headquartered, auditable, self‑hostable model from a $14B‑valued lab matters in procurement decisions even if it doesn’t top performance charts.
Bottom line for the crypto ecosystem
For crypto teams, DAOs, and exchanges weighing AI tooling, the tradeoffs are familiar: raw benchmark performance versus auditability, legal exposure, and the ability to self‑host. Mistral’s Medium 3.5 is not the cheapest or the top performer—but it keeps European open‑weight LLMs in the conversation, and that jurisdictional, legal, and hosting flexibility can be decisive for firms that can’t—or won’t—route sensitive workloads through foreign cloud providers.
Read more AI-generated news on: undefined/news
