AI Potluck
Model components / Fine-tuning code

Mistral Fine-Tuning API

Mistral AI

Managed LoRA fine-tuning of Mistral 7B, Mistral Small, and other Mistral models via La Plateforme, plus an open source companion (mistral-finetune) for self-hosted training. Picked when teams want Mistral's European data residency story or open-weight portability between managed and self-hosted training. API priced at ~$2/M training tokens; the only managed offering for Mistral's commercial weights.

Mistral serverless fine-tuning on La Plateforme - confirmed live/available in 2026 (not deprecated; only the legacy docs guide moved to /resources/deprecated/). Methods: LoRA-adapter SFT (text & vision) + Classifier Factory. Tunable: Mistral Small / Medium (originally 7B + Small). $4 min job fee + $2/mo storage per model. Proprietary managed service. (Separate mistral-finetune OSS SDK exists for self-hosting - not this product.)

Openness

1 high confidence
1.0
source
closed
training-pipeline
closed
runs-on
La-Plateforme-only
license
Proprietary(managed service, per-token + storage pricing)
base weights
Mistral-models(open-weight variants exist)

The hosted FT API on La Plateforme is a proprietary managed service (LoRA adapters under the hood); no source for the service. Mistral separately ships an OSS mistral-finetune SDK for self-hosting, but that is a distinct product from this managed API. Verified 2026-06-04: the FT API remains a live production service on La Plateforme (changelog + platform docs); only the legacy docs guide was relocated to /resources/deprecated/, which carries a deprecation banner but does not retire the API.

  • https://mistral.ai/news/customization recorded 2026-06-04

    serverless FT on La Plateforme; LoRA adapters under the hood; managed proprietary service; separate OSS mistral-finetune SDK

  • https://docs.mistral.ai/resources/deprecated/finetuning recorded 2026-06-04

    legacy fine-tuning DOCS guide moved to /resources/deprecated/ (Classifier Factory + text/vision SFT documented); corroborates that only the docs guide moved while the managed FT API on La Plateforme remains a live production service

Adoption

not assessed

No disclosed standalone usage figure for the Mistral fine-tuning API located this run (jobs run / customers tuning). La Plateforme is a widely used developer platform, but no honest per-SKU FT signal was found; declining to assign a level.

Capability

3 medium confidence
3.0

Narrower managed method set than the SFT+DPO+RFT services - LoRA SFT + classifier factory only, over Mistral's own models, black-box. Scored 3: a real, capable managed LoRA-FT service but below the preference/RL-capable OpenAI/Azure/Together/Fireworks offerings (C4) and far below the OSS scale definers.

Unchanged since 2026-06-09 (last edited, not re-checked)