This update brings computer use (beta), letting local models see your screen and interact with apps, plus MCP integrations for Unity, Unreal Engine, Blender and Roblox.
MiniMax-H3-Image now supports GGUF, FP8 and INT8, with LoRA training. We've also introduced a new model picker, redesigned Connections and improved project folders.
Highlights
- Engram: Change and swap engrams for Qwen Flash and DeepSeek.
- Computer use (beta): Local vision models and connected providers can now see your screen and interact with apps. Off by default, with permission controls.
- Game engines + creative apps: Connect Unreal Engine, Unity, Godot and Roblox through MCP, alongside Blender.
- MiniMax-H3-Image: GGUF support for CPU, Mac and AMD, plus FP8 and INT8 weights for NVIDIA. LoRA training is also supported.
- New model picker: Switch models directly from the composer, with Favorites, On Device, Discover and Connected tabs.
- Revamped Connections: Live connection status, a redesigned provider gallery and 9 new providers.
- Auto compaction: Long conversations are automatically summarized, with configurable limits for each connection.
- New project folders: Browse project files, search folders and view Git changes from a redesigned Files tab.
Computer Use (Beta)
Models can now see your screen and control the mouse and keyboard across macOS, Windows and Linux.
- Supports local GGUF, safetensors and MLX vision models, plus connected providers.
- Requires approval before the first action, switching apps and potentially risky actions.
- Blocked apps remain inaccessible, and you can take control at any time.
Computer use guide: https://unsloth.ai/docs/basics/computer-use
Game Engines + Creative Apps
Unity, Unreal Engine, Godot, Figma and Roblox are now available in the MCP menu alongside Blender.
Connect your apps to let models build scenes, create games and interact with creative projects. Figma is read-only.
Game engines guide: https://unsloth.ai/docs/basics/mcp/mcp-engines
MiniMax-H3-Image
MiniMax-H3-Image now supports more hardware and faster inference.
- GGUF runs through stable-diffusion.cpp on CPU, Macs, smaller GPUs and AMD.
- FP8 and INT8 weights are available for NVIDIA GPUs.
- Automatic precision selects INT8, FP8 or BF16 based on hardware.
- Train and load custom LoRAs.
A 2048px image decodes in 0.28s instead of 2.82s on an A100 80GB.
GGUF weights: https://huggingface.co/unsloth/MiniMax-H3-Image-GGUF
FP8 / INT8 weights: https://huggingface.co/unsloth/MiniMax-H3-Image-FP8
Model Picker + Projects
The new composer model picker replaces the Thinking pill, bringing model selection and reasoning settings into one place.
Projects also get redesigned folders, a Files tab, Git change tracking and quicker folder navigation.
Connections + Compaction
Connections now display live statuses, including Connected, Key rejected, No credit and Rate limited.
The provider gallery adds xAI, NVIDIA NIM, Perplexity, Cohere, Vercel AI Gateway, Azure OpenAI, Amazon Bedrock, Google Vertex AI and SGLang.
Long conversations on connected providers are now automatically summarized instead of losing older messages. You can configure compaction limits per connection, and /compact now works with cloud models.
Engram Tables
Qwen3.8-Flash-Next and DeepSeek-V4.1 GGUFs keep their Engram table in its own file, so you can now run any quant with a smaller or lossless table. The Qwen3.8-Flash-Next and DeepSeek-V4.1 guides list each table's size.
- Click a model's gear icon in the model picker, turn on Advanced settings, then pick a table under Memory > Engram Table.
- The table you pick downloads on the next load, and the quant's own table is skipped if it isn't downloaded yet.
- From the terminal, add
--engram, likeunsloth run --model unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL --engram Q8_0.
Training + Performance
- Gemma 4 12B training is 1.4–2.6x faster with new attention kernels.
- DeepSeek-V4.1 adds native LoRA, QLoRA and distributed training support.
- SFT now packs text datasets by default on epoch-based runs.
- LoRA training is available for FLUX.3 Klein and Qwen-Image-2.1.
- New INT8 attention kernels improve FLUX.3 Klein rendering by 15–23% on supported GPUs.
What's Changed
This release also includes extensive improvements across Unsloth Studio and training:
- Improved model loading, image and video generation, and LoRA workflows.
- Faster INT8 attention and diffusion kernels.
- Expanded distributed training support, including FSDP2 and tensor parallelism.
- Improved AMD GPU and NPU support.
- Better browser, chat, project and connection reliability.
- Numerous bug fixes, performance optimizations and compatibility updates.
Thanks to everyone who contributed to this release!
New Contributors
- @jaeminbird made their first contribution in #17285
- @gabrielkmbo made their first contribution in #17328
- @csun365 made their first contribution in #17121
- @bezabyte made their first contribution in #17648
- @RioPlay made their first contribution in #17645
- @Lluczel made their first contribution in #17731
- @bjornts made their first contribution in #13929
- @ellaslan11 made their first contribution in #17938
- @zubincarvalho made their first contribution in #17933
- @Tofuwang45 made their first contribution in #17879
Full Changelog: v0.1.917-beta...v0.1.918-beta