Mistral: Our first AI endpoints are available in early access

Original Article ↗ Hacker News Discussion ↗

Mistral AI’s early-access API endpoints, including the closed-source “mistral-medium” model and open-weight smaller variants, are drawing attention for benchmark scores that reportedly surpass GPT‑3.5 and competitive pricing against OpenAI and Anthropic. Commenters weigh the value of paying for a hosted service when strong open models can be self‑deployed, citing reliability, portability, and protection against vendor lock-in as key reasons to still use third‑party APIs. There is also active debate over real-world performance vs. benchmarks, hardware and energy costs of large-scale inference, context window limitations, and how EU regulation and open-source exemptions might shape Mistral’s long-term position.

Model lineup and capabilities

  • New endpoints: Mistral-tiny (Mistral 7B Instruct v0.2), Mistral-small (open weights), Mistral-medium (closed, prototype), plus an embedding model.
  • Mistral-medium is claimed to beat GPT‑3.5 and even edge GPT‑4 on some benchmarks (e.g. Winogrande), and significantly outperform Gemini Pro; some see this as “extremely impressive” if it holds in real use.
  • Others caution that benchmarks are easy to game, are not apples‑to‑apples, and that real-world performance and safety-tuned public versions may differ.
  • Mixtral 8x7B MoE is highlighted as performing near 70B dense models at roughly 14B inference cost; some report it doing very well on practical tasks.

Pricing and economic viability

  • Posted prices per 1M output tokens place Mistral-medium slightly above GPT‑3.5 and far below GPT‑4; some infer Mistral believes it’s clearly better than GPT‑3.5.
  • Others argue GPT‑3.5 is subsidized, or that early-access pricing may not reflect eventual reality.
  • There is debate over whether cheaper but slightly worse models can win against OpenAI, and how much reliability and openness matter compared to raw quality.

Open-source vs hosted APIs

  • Mistral’s open-weight models are seen as a major strategic advantage: users can start with hosted APIs and later self-host or move providers, reducing vendor lock-in and “rug-pull” risk.
  • Some argue that if models are open, many providers can host them cheaply, eroding Mistral’s API margins.
  • Open models are also seen as insurance against model deprecations and behavior changes.

Context window and architecture

  • Mixtral’s 32k context with sliding-window attention draws mixed reactions.
  • One side reports that for tasks needing precise verbatim recall (code reuse, business documents), sliding-window models effectively behave like ~8k context and fail; they work better for gist-level reasoning.
  • Others note that hidden states retain information and dispute that it’s so limiting; impact is described as “unclear” and use‑case dependent.

Hardware, energy, and local inference

  • Several comments estimate power and cost per million tokens on GPUs (4090 vs A100/H100), with batching and data-center energy prices significantly affecting economics.
  • There is speculation about running MoE models locally (even on phones) via compression/quantization, with disagreement about practicality due to latency and power draw.

Ecosystem, regulation, and competition

  • Many see Mistral as a surprisingly strong small player and a serious threat to incumbents for non‑GPT‑4 use cases.
  • Some worry EU AI regulation and being based in Europe could be a long‑term handicap; others argue exemptions for open models and “building for regulation” may become an advantage.
  • Google’s perceived strategic confusion in AI and search monetization is contrasted with Mistral’s focused execution, though some caution against writing Google off.