Samba-1 Turbo 25.4.1-MP2

Release version: 25.4.1-MP2 | Release date: 07/01/2025

The Samba-1 Turbo 25.4.1-MP2 (Model Pack 2) release focuses on expanding integration capabilities and improving the user experience.

New and updated model versions

New models

  • Aurora GPT

  • Llama-4-Maverick-17B-128E-Instruct

  • Llama-4-Scout-17B-16E-Instruct

  • Qwen3-32B

  • Multilingual E5 Large

  • Multilingual E5 Large Instruct

  • DeepSeek-R1-16384

  • DeepSeek-V3-16384

  • DeepSeek-R1-32768

  • DeepSeek-V3-32768

  • DeepSeek-R1-4096

  • DeepSeek-V3-4096

  • DeepSeek-R1-8192

  • DeepSeek-V3-8192

  • Llama-4-Maverick-17B-128E-Instruct-131072

  • Llama-4-Maverick-17B-128E-Instruct-16384

  • Llama-4-Maverick-17B-128E-Instruct-65536

  • Llama-4-Maverick-17B-128E-Instruct-8192

Updated models

This release includes updated versions of the following model(s):

  • DeepSeek-R1

  • DeepSeek-V3

Quality improvements

  • E5 Instruct models are now available via the OpenAI-compatible API for seamless integration.

Known issues

  • Playground is currently enabled but broken for models that are not designed to support it (e.g., Stormer).

  • Embedding models in CoEs are incorrectly selectable in the Playground UI.

  • CoE bundles with different models using the same PEFs may not function as expected.

    • A fix is planned for 25.7.1-MP1.

  • LLaMA 4 supports text-only mode in the Playground; vision support is API-only.

  • LLaMA 4 Maverick / DeepSeek models may appear selectable for flexible CoEs, but these CoEs are likely to fail.

  • Modelcards for custom prebuilt LLaMA 4 / DeepSeek CoEs incorrectly list all configs as available.