FlowtandBriefAI · Software · ResearchMeet Flowtand

Mistral Opens Beta of La Plateforme With Three Chat Endpoints and an Embedding API

Announcement

Mistral AI

Browse Industry topic page →

Mistral announced beta access to La Plateforme, offering three chat endpoints (mistral-tiny, mistral-small, mistral-medium) with differing performance/price tradeoffs plus mistral-embed, a 1024-dimension embedding endpoint. Mistral reports MT-Bench scores of 7.6, 8.3 and 8.6 respectively and says the API follows a familiar chat interface, with Python and JavaScript clients.

What Mistral announced

Mistral opened beta access to La Plateforme, its first platform services. The offering consists of three chat endpoints for generating text from instructions and one embedding endpoint, each with a different performance/price tradeoff. Mistral-tiny serves Mistral 7B Instruct v0.2 and, per Mistral, works only in English and scores 7.6 on MT-Bench. Mistral-small serves Mixtral 8x7B, which Mistral says handles English, French, Italian, German, Spanish and code, scoring 8.3. Mistral-medium serves a prototype model scoring 8.6 and described as among the top serviced models on standard benchmarks. The embedding endpoint, mistral-embed, produces 1024-dimensional embeddings and was designed with retrieval in mind, with Mistral reporting a 55.26 retrieval score on MTEB. Mistral says the API follows a popular chat interface specification, provides Python and JavaScript client libraries, and allows a system prompt to raise moderation on outputs. All benchmark numbers are Mistral's own reported figures.

Who might use it

A backend engineer building a multilingual support assistant could trial mistral-small for conversations that mix European languages and code snippets, while routing simpler English-only requests to mistral-tiny. A search or RAG team could pair mistral-embed with a chat endpoint to retrieve passages before generating answers. Because all three endpoints share a chat-style interface and system prompts, a team could try keeping one integration layer and swapping endpoints by configuration rather than rewriting code. These are conditional options suggested by Mistral's described endpoints, not claims that any specific integration works.

A concrete trial

A team could start by registering for beta access, then run a small evaluation set of its own prompts and retrieval queries through each endpoint to compare quality, latency and cost against its current baseline system. For retrieval, measure whether the intended passages appear in top results using the team's own labelled queries. Track failure modes such as non-English prompts on mistral-tiny, which Mistral says is English-only, or output quality changes between endpoints. Because medium serves a prototype and access is beta, keep a human review step for high-stakes outputs and do not assume stable behaviour. Mistral warns of rough edges during the ramp to self-served availability.

Trade-offs and unknowns

Choosing among endpoints is a trade-off between reported quality and cost, but the excerpt does not state actual prices or rate limits, so budgeting requires checking the platform directly. The MT-Bench and MTEB scores are publisher-reported and not independently verified. Mistral-medium is a prototype, which implies behaviour may change, and the whole service is beta with capacity being ramped progressively. Teams with strict data-handling or compliance requirements should not treat this excerpt as evidence of those guarantees. A team might avoid migrating production workloads until general availability is clearer.

Context and limitations

  • Based on a short source excerpt; the full article was not analyzed.
  • Benchmark figures (MT-Bench, MTEB retrieval) are Mistral's own reported numbers, not independent tests.
  • Access is beta: Mistral says to expect rough edges as it ramps to general availability.
  • Based on bounded source text, not independent testing. Practical suggestions are editorial proposals.