Mosaic AI Model Training

source

closed

training-code

closed

runs-on

Databricks-workspace-only(AWS regions)

license

Proprietary(managed service)

base-weights

Llama-Community-License(non-OSI)

A proprietary managed fine-tuning service that runs only inside a Databricks workspace, with no source published. Databricks' separate LLM Foundry and Composer libraries are open source, but they are not this product.

• https://docs.databricks.com/aws/en/large-language-models/foundation-model-training/ recorded 2026-08-09

  Foundation Model Fine-tuning requires a Databricks workspace in AWS us-east-1 or us-west-2 and is used only through the databricks_genai SDK or its UI; the page publishes no repository or self-hosted build of the training engine, and the only license text on the page covers the base Llama models being fine-tuned, not the service itself.

• https://www.databricks.com/product/machine-learning/mosaic-ai-training recorded 2026-08-09

  Databricks markets this capability as Databricks Model Training, describing it only as a way to fine-tune an open source LLM on Databricks; the page offers no download, repository, or self-host option for the training service itself, and never uses the name Mosaic AI Model Training.

No usage figure is disclosed for the fine-tuning feature -- no jobs run, no count of customers tuning. Databricks as a whole is widely used, but nothing published is specific to this capability, so no level is assigned. The feature is also being sunset, with removal set for Aug 14 2026, which further depresses any forward adoption.

• https://docs.databricks.com/aws/en/large-language-models/foundation-model-training/ recorded 2026-08-09

  The Foundation Model Fine-tuning page discloses no usage, job-count, or customer figures for the feature anywhere in the article; the only forward-looking statement is the admonition that it is deprecated and scheduled for removal on August 14, 2026.

• https://www.databricks.com/product/machine-learning/mosaic-ai-training recorded 2026-08-09

  The Databricks Model Training product page carries no customer counts, job counts, or usage figures specific to this fine-tuning capability, only generic cost-comparison marketing language (e.g. lower cost than proprietary LLMs).

A narrow managed method set: chat-completion and instruction fine-tuning plus continued pretraining, with no preference or RL methods, no exposed LoRA or QLoRA, no parallelism, precision or scale controls, and base models limited to the Meta Llama family. That places it below the managed offerings from OpenAI, Vertex and Together, which combine supervised, preference and reinforcement tuning and reach 100B+ models, and far below the open-source libraries that define scale. The feature is also being deprecated, which caps any further investment in it.

• https://docs.databricks.com/aws/en/large-language-models/foundation-model-training/ recorded 2026-08-09

  Supported tasks lists exactly three methods, chat completion, instruction fine-tuning, and continued pre-training, with no LoRA, QLoRA, DPO, RLHF or other preference-tuning options mentioned; Supported models lists only Meta Llama 3.1, 3.2 and 3.3 checkpoints (70B/8B and 1B/3B variants), matching the recorded feature-matrix value exactly.