Model components / Fine-tuning code

Mosaic AI Model Training

Databricks' managed full-fine-tune and continued-pretraining service for open-weight models (Llama 3, Mistral, DBRX) built on the proprietary MosaicML training stack, claimed up to 2x faster than naive open source. Picked when training data lives in Unity Catalog and teams want Lakehouse-native fine-tuning with full weight ownership (resulting model registers to UC). Distinct from competitors in that customers get the actual weight checkpoint, not just a hosted endpoint.

Databricks Foundation Model Fine-Tuning (now Databricks Model Training), accessed via the databricks_genai SDK. Confirmed live in docs June 2026 BUT marked deprecated with scheduled removal Aug 14 2026. Methods: chat-completion FT, instruction FT, continued pre-training. Base models: Meta Llama 3.1/3.2/3.3 family only. Proprietary managed Databricks service.

Openness

1 high confidence

1.0

source: closed
training-code: closed
runs-on: Databricks-workspace-only(AWS regions)
license: Proprietary(managed service)
base weights: Llama-Community-License(non-OSI)

Proprietary managed FT service running only inside a Databricks workspace; no source. (Databricks' separate LLM Foundry / Composer libs are OSS but are not this product.)

https://docs.databricks.com/aws/en/large-language-models/foundation-model-training/ recorded 2026-06-04
managed FT service via databricks_genai SDK; deprecated, removal Aug 14 2026; Llama-only base models

Adoption

not assessed

No disclosed usage figure for the FMFT feature located this run (jobs run / customers fine-tuning). Databricks overall is widely used, but no honest signal specific to this FT capability was found; declining to assign a level. Feature is also being sunset (removal Aug 14 2026), which further depresses any forward adoption.

https://docs.databricks.com/aws/en/large-language-models/foundation-model-training/ recorded 2026-06-04
managed FT feature; deprecation notice; no usage figures disclosed

Capability

2 medium confidence

2.0

Narrow managed method set (SFT + continued pretraining only, no preference/RL methods) over a single base-model family, and a black-box service. Scored 2: below the OpenAI/Vertex/Together managed offerings (SFT+DPO+RFT / 100B+ models) and far below the OSS scale definers. Being deprecated also caps forward capability investment.

https://docs.databricks.com/aws/en/large-language-models/foundation-model-training/ recorded 2026-06-04
chat-completion FT, instruction FT, continued pre-training; Llama-only

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