Mistral Medium 3.5

Original Article ↗ Hacker News Discussion ↗

Mistral Medium 3.5, a 128B-parameter dense, text‑only model from French startup Mistral, is drawing mixed reactions as an open‑weights alternative to US and Chinese “frontier” AI systems. Commenters note that while it appears competitive with strong mid‑tier models and is attractive for on‑premise and sovereignty‑focused deployments, it lags leading proprietary and newer Chinese models on capability, speed, and price, especially given its heavy hardware requirements. The release is seen as part of a broader strategy where smaller labs aim for “good enough” models optimized for cost and local control, even as many users feel the performance gap to top agents from OpenAI, Anthropic, and DeepSeek still meaningfully impacts productivity.

Perceived performance vs frontier models

  • Many commenters see Mistral Medium 3.5 as “okay but not exceptional” compared to frontier models like GPT‑5.5, Claude Sonnet/Opus, and top Chinese models (DeepSeek, GLM, Qwen, Kimi).
  • Some argue that for typical coding and chat tasks, differences vs frontier models are small; others say for complex agentic workflows, the gap is now “enormous” and materially impacts productivity.
  • Several note that smaller Chinese and Google models (e.g., Qwen 3.6 27B, Gemma 4 26–31B) match or beat it despite being much smaller.

Benchmarks and evaluation concerns

  • Launch blog leans on SWE‑Bench Verified, which some distrust due to alleged contamination and past disputes between labs.
  • Multiple users say the model performs poorly on SVG/HTML/JS generation, especially compared to Gemma and Kimi; others downplay SVG quality as a meaningful metric.
  • There’s skepticism about claims that it “beats Sonnet,” with people reporting open‑weights generally lag Sonnet in practical agent tasks despite benchmark wins.

Pricing and competitiveness

  • The model is viewed as expensive: significantly more than Mistral Large and Chinese competitors, and more than Anthropic’s Haiku / some Sonnet‑tier options.
  • Some praise earlier Mistral models (Large, Small 4) as Pareto‑competitive (80–90% of frontier quality at much lower cost); this release is seen as less clearly on that frontier.

Open‑weight, dense design and local deployment

  • Medium 3.5 is a 128B dense, open‑weight, 256k‑context model (~140 GB full; ~70–80 GB at Q4 quant).
  • Enthusiasts like that it can, in principle, run locally on high‑end Macs or multi‑GPU rigs and offers sovereignty vs US/Chinese clouds.
  • Others point out the physics: dense 128B on consumer hardware yields very low tokens/sec; MoE alternatives (e.g., DeepSeek V4 Flash, Qwen 35B A3B) give higher effective capability per byte and far better speeds.
  • Debate over why Mistral chose a large dense model given its own earlier MoE success; some see this as a strategic misstep.

Use cases, tools, and product experience

  • Positive experiences with older Mistral models for text transformation, document analysis, and on‑prem enterprise deployments.
  • Concerns that the new Medium’s higher price may foreshadow deprecation of cheaper Large.
  • Mixed feedback on Mistral Vibe (coding agent) and CLI: some like the concept; others report bugs, instability, strict CSP preventing easy JS demos, and weak coding/tool behavior versus Claude Code, Codex, or OpenCode.

Geopolitics and ecosystem

  • Strong interest in a credible non‑US, non‑Chinese option for regulatory, political, and “data sovereignty” reasons.
  • Some worry Europe chronically underinvests and lags the US/China, while others argue efficient training and open‑weights can still make Mistral strategically important even if it is not strictly SOTA.