Intel / repo

unsloth v0.1.917-beta — Dynamic v3.1 GGUFs + Performance Fixes

Source: github.com
Author: shimmyshimmer
Date: Published Nov 11, 2026

Context

Unsloth's latest Studio release lets a Mac user run Anthropic's Claude Code or OpenAI's Codex against a model running on their own machine through Apple's MLX framework. Unsloth Studio is a local app for running and open models, and its `unsloth start claude` and `unsloth start codex` commands launch a coding that talks to the model Studio is serving instead of a cloud service. Until this release those commands refused any model that was not in GGUF, the file format used by the llama.cpp inference server.

MLX is Apple's machine-learning library for its own chips, and models converted to it run natively on Apple Silicon Macs. According to the pull request, Studio already answered MLX requests on the API routes the two agents call; what blocked them was a check in `unsloth start`. That check now accepts an MLX model when the server reports an Apple Silicon Mac, while ordinary Transformers checkpoints are still refused. The same change fixed a bug in which Qwen3's chat template made Claude Code sessions fail with a server error after the agent compacted its history, and the release notes list a separate fix so an MLX model that runs out of returns an error instead of an empty reply.

The practical result is one more way to give these agents a local model on a Mac without converting it to GGUF first. The release does not measure how well either agent performs on MLX models, so the change is about compatibility, not quality.

Key takeaways · AI-distilled

Unsloth says model.generate() now runs 2-8x faster by default on T4 GPUs across its FastModel models.
INT8 and FP8 image models now load on older NVIDIA GPUs including the T4, and the FP8 build of LTX-2.5 now runs on 24 GB GPUs, so these image and video models no longer require newer cards.
Claude API models called without an explicit token limit no longer stop at 1,024 tokens, and MLX models now return a context-length error instead of an empty response when a prompt is too long.
On the training side, Ulysses context parallelism now supports Kimi-Linear, DeepSeek, MiniMax and Gemma 4 text models, and video columns are now handled correctly when training vision models.
In the Studio chat app, performance no longer degrades as conversations grow, and read-aloud starts after the first sentence instead of waiting for the whole response.

Terms in this piece · Glossary

fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
streaming — Sending a model's response token by token as it is generated, so the reader sees text immediately instead of waiting for the whole answer.

Read the source github.com