Mistral Fine-Tuning API

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)

Mistral's hosted fine-tuning API is a proprietary managed service running LoRA adapters, with no source published; a separate mistral-finetune SDK covers self-hosting but is a different product. Mistral's docs now file the API under deprecated, legacy features, though current pricing still bills its Classifier tier, so the service remains closed with no self-hostable build.

• https://docs.mistral.ai/resources/deprecated/finetuning recorded 2026-09-23

  Nav files this page under 'Deprecated features' as 'Fine-tuning API (legacy)'; the page banner reads 'This feature is deprecated and is no longer actively supported.' Describes the service as 'Text and vision general fine-tuning via SFT' plus 'Classifier Factory', billed with 'a minimum fee per fine-tuning job of $4' and a monthly storage fee; no repository or self-hosted build path is offered for this managed feature.

• https://legal.mistral.ai/terms recorded 2026-09-23

  Indexes Mistral's legal documents, listing a 'Commercial Terms of Service' for commercial/API customers and a 'License Notice' for everyone; no open source license is offered for the hosted platform itself, consistent with a proprietary managed service.

• https://mistral.ai/news/customization recorded 2026-09-23

  States it will 'leverage our proprietary fine-tuning techniques with our managed fine-tuning services' for the serverless offering, and separately introduces the open source mistral-finetune SDK as the self-hosted alternative; no repository is offered for the managed service itself.

• https://mistral.ai/pricing/api/ recorded 2026-09-23

  Its 'Browse our APIs' catalog lists 'Classifier API model 3B' and 'Classifier API model 8B' fine-tuning, each billed 'minimum fee per fine-tuning job of $4' plus a per-token training cost and a monthly per-model storage cost; this is Mistral's own commercial price list, not a license grant, and no separate catalog entry prices general text/vision SFT.

Mistral publishes no usage figures for the fine-tuning API: no jobs run, customers tuning, models fine-tuned or developer count. The wider Mistral platform is widely used, but nothing published is specific to fine-tuning, so adoption is left unscored rather than inferred from the platform around it.

• https://docs.mistral.ai/resources/deprecated/finetuning recorded 2026-09-28

  The overview page carries the banner "This feature is deprecated and is no longer actively supported." and lists "via SFT: supervised fine-tuning, the most common fine-tuning method" alongside Classifier Factory. Searched the body for DPO, reinforcement and LoRA: none found. Searched for downloads and usage figures: none.

• https://mistral.ai/news/customization recorded 2026-09-28

  The launch post presents the managed service on la Plateforme and says "We use LoRA adapters under the hood to prevent forgetting base model knowledge and allow for efficient serving." Searched the body for downloads, jobs run, active users and customer counts for the fine-tuning service: no usage figure found.

Mistral's fine-tuning API offers supervised fine-tuning for text and vision, plus a Classifier Factory, on Mistral's own models, with only training steps and learning rate exposed as hyperparameters and no full fine-tuning, DPO, or reinforcement fine-tuning documented. That is narrower than the preference- and RL-capable services from OpenAI, Azure, Together and Fireworks.

• https://docs.mistral.ai/resources/deprecated/finetuning recorded 2026-09-28

  The overview page carries the banner "This feature is deprecated and is no longer actively supported." and lists "via SFT: supervised fine-tuning, the most common fine-tuning method" alongside Classifier Factory. Searched the body for DPO, reinforcement and LoRA: none found. Searched for downloads and usage figures: none.

• https://docs.mistral.ai/resources/deprecated/finetuning/text_vision_finetuning recorded 2026-09-28

  The Text & Vision fine-tuning guide carries the banner "This feature is deprecated and is no longer actively supported." and lists the tunable models "open-mistral-7b", "mistral-small-latest", "codestral-latest", "open-mistral-nemo", "mistral-large-latest", "ministral-8b-latest", "ministral-3b-latest" and "pixtral-12b-latest", with "training_steps" and "learning_rate" as the adjustable hyperparameters; its FAQ says "For LoRA fine-tuning, we recommend 1e-4 (default) or 1e-5." Searched the body for DPO, preference (as a tuning method), reinforcement and full fine-tuning: none.

• https://mistral.ai/pricing/api/ recorded 2026-09-28

  The API price list itemizes fine-tuning only for the classifier entries, for example "Fine-tune Ministral 3B for classification tasks, like moderation, sentiment analysis, fraud detection, and more." with "minimum fee per fine-tuning job of $4". Searched the body for DPO, reinforcement and full fine-tuning: none.