Unsloth AI · 2026-04-23 · major
Unsloth Studio v0.1.37 — New UI Redesign, Qwen3.6 Support, Preserve Thinking Mode
Open-source local LLM fine-tuning web UI ships major redesign with Qwen3.6 support, a Preserve Thinking toggle for reasoning models, developer role support for Codex/OpenCode, and improved tool-call parsing. 64k stars.
Unsloth Studio's April redesign adds Qwen3.6 support, Preserve Thinking toggle, and developer role for agentic coding tools.
Key specs
| GitHub stars | 63,700 |
|---|---|
| Models supported | 500 |
What is it?
Unsloth Studio is an Apache 2.0 open-source web UI for training and running open AI models locally on Mac, Windows, and Linux. Version 0.1.37-beta landed April 23, 2026 with a complete UI redesign and support for the newly-released Qwen3.6 model family.
How does it work?
The UI redesign unifies the fine-tuning, inference, and model export workflows into a cleaner interface. New Qwen3.6 support includes Dynamic GGUF quantization for running 27B and 35B-A3B models locally. The Preserve Thinking toggle retains the reasoning trace from the previous turn for models like Qwen3.6 that support extended thinking, trading tokens for improved reasoning continuity. Developer role support adds compatibility with Codex and OpenCode agentic coding tools. Improved tool-call parsing fixes nested object handling that caused failures in multi-tool agent flows.
Why does it matter?
Qwen3.6-27B runs in 18GB RAM via Unsloth's Dynamic GGUFs and outperforms Qwen3.5-397B on coding benchmarks — being able to fine-tune and run it locally democratizes access to frontier-quality coding models for independent developers.
Who is it for?
Developers and researchers who want local fine-tuning and inference of 2026's best open-weight models without cloud costs.
Try it
pip install unsloth && unsloth-studio