Skip to content
AIBRAINWORX / OpenSourceCoder.comLOCAL-FIRST. BUILT IN THE OPEN.
Developer & Agent Tools / Inference & Quantization

Unsloth v0.1.919-beta Adds Local Computer Use, MiniMax-H3-Image Quantization, and Game Engine MCPs

Unsloth ships v0.1.919-beta with opt-in desktop Computer Use, GGUF/FP8/INT8 MiniMax-H3-Image inference, decoupled Engram tables, and MCP connectors for five creative engines.

Listen to this articleBrowser Voice
Ready
Official company logo or product image for unsloth-v0-1-919-beta-computer-use-minimax-h3-image-mcp
Official image or organization logo from github.com; identification only, no endorsement.
Event / context: 2026-11-17. Source check: October 28, 2026.

Unsloth has published release v0.1.919-beta, a wide-ranging update to its open-source model training and local inference suite authored by maintainer Daniel Han alongside 33 contributors. Released on November 17, 2026 under verified commit 38be5c7, the update advances Unsloth Studio beyond fine-tuning scripts into an integrated local-and-cloud development environment. Headlining the release are opt-in desktop Computer Use for vision models, quantized inference and LoRA support for MiniMax-H3-Image, decoupled Engram memory tables for flagship open-weight LLMs, a unified composer model picker, local page translation, and Model Context Protocol (MCP) connectors for major game and design engines.

Scope Across Local Agents, Workspaces, and Cloud Providers

The operational scope of v0.1.919-beta spans desktop automation, local text processing, workspace navigation, and remote API orchestration. The new Computer Use (Beta) feature lets local vision models in GGUF, safetensors, and MLX formats—as well as connected cloud providers—view the screen and control mouse and keyboard input across macOS, Windows, and Linux. Alongside desktop control, Unsloth added Translate Page to translate text locally into any language and overhauled project workspaces with a redesigned Files tab supporting file browsing, folder search, and Git change tracking. For hybrid workflows, the revamped Connections panel displays live connection health (Connected, Key rejected, No credit, and Rate limited) and expands the provider gallery with nine additions: xAI, NVIDIA NIM, Perplexity, Cohere, Vercel AI Gateway, Azure OpenAI, Amazon Bedrock, Google Vertex AI, and SGLang.

Precision Formats, Engram Tables, and Kernel Metrics

At the inference and training layer, Unsloth targets broader cross-vendor hardware reach and lower memory footprints. For MiniMax-H3-Image, GGUF weights now run via stable-diffusion.cpp on CPUs, Macs, AMD hardware, and smaller GPUs, while NVIDIA GPUs gain dedicated FP8 and INT8 weights. Automatic precision selection chooses INT8, FP8, or BF16 based on detected hardware, and custom LoRA training and loading are supported. According to Unsloth's release notes, a 2048px image decodes in 0.28 seconds instead of 2.82 seconds on an NVIDIA A100 80GB GPU. For text models, Qwen3.8-Flash-Next and DeepSeek-V4.1 GGUF packages separate their Engram table into an independent file, allowing users to pair any quantization with a smaller or lossless table via the model picker under Advanced settings > Memory > Engram Table or on the CLI with flags such as unsloth run --model unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL --engram Q8_0. On the training side, Unsloth reports 1.4x to 2.6x faster Gemma 4 12B training using new attention kernels, native LoRA, QLoRA, and distributed training (including FSDP2 and tensor parallelism) for DeepSeek-V4.1, default text-dataset packing for epoch-based SFT runs, LoRA training for FLUX.3 Klein and Qwen-Image-2.1, and 15% to 23% faster FLUX.3 Klein rendering on supported GPUs via INT8 attention kernels.

Developer Workflows, MCP Engines, and Context Compaction

For day-to-day engineering and creative production, the release consolidates fragmented controls into single-surface workflows. The new composer model picker replaces the older Thinking pill, placing model selection and reasoning settings into one interface organized across Favorites, On Device, Discover, and Connected tabs. Beyond model execution, the MCP integration menu now connects Unity, Unreal Engine, Godot, Roblox, and Figma alongside existing Blender support, enabling models to construct scenes, assist with game workflows, and inspect design projects. For lengthy cloud-backed sessions, Unsloth added automatic conversation compaction that summarizes older messages instead of truncating them. Developers can configure compaction thresholds per connection or manually trigger /compact on cloud models, preserving task context alongside the workspace's new Git diff tracking and improved AMD GPU and NPU support.

Beta Guardrails, Hardware Caveats, and Licensing Boundaries

Several practical and verification limits apply to v0.1.919-beta. Computer Use is explicitly labeled beta and remains off by default; while Unsloth requires user approval before the first action, when switching apps, and before potentially risky actions—and enforces blocked-app restrictions with instant user takeover—granting OS-level input control to vision models still warrants strict sandboxing. Within the expanded MCP suite, Figma integration is strictly read-only, preventing direct canvas modifications. Furthermore, all reported performance gains, including the 0.28-second A100 image decode, Gemma 4 12B kernel speedups, and FLUX.3 Klein INT8 improvements, are self-reported in Unsloth's release notes without independent benchmark verification, and INT8 kernel acceleration applies only to supported GPUs. Finally, the release announcement does not enumerate individual weight licenses; developers should review upstream Hugging Face model cards for MiniMax-H3-Image, Qwen3.8-Flash-Next, and DeepSeek-V4.1, as well as third-party API privacy and billing terms when using connected cloud providers and automated context compaction.

EXPLORE THE SOURCE

GitHub repositories

Automated Vertex AI (gemini-4-hydrogen) primary-source report; reviewed 2026-11-18.

Follow your curiosity.

Pass the signal.