Unsloth adds local DeepSeek and FLUX models

Unsloth’s October 19, 2026 release adds local support for DeepSeek-V4.1-Flash-1017 and FLUX.3 Image Klein. The DeepSeek model has 552B parameters (16B active) and runs on ~125 GB at 1-bit with its Engram table stored on SSD. FLUX.3 Image Klein runs on 24 GB and supports editing with up to 10 reference images.

The update also brings images and videos into chat, allowing local chat models to use an image or video model loaded alongside them. A new MCP setup command connects Claude Code, Codex or Cursor to Unsloth. Training additions include multi-GPU support through FSDP2, currently in alpha, and a token preview showing what a model will learn before training begins.

Recent releases also introduced browser control, enabling models to open pages, read, click, type and scroll while users watch. Other additions include text-diffusion model training and chat, improved prompt caching for coding agents, and support for NVIDIA Nemotron 3.5 models and Meta Muse Spark Open. The broader desktop platform supports local inference and training across Mac, Windows and Linux, with ongoing improvements to memory use, hardware compatibility and sandboxed tool execution.