Discussion

Unsloth’s beta release adds local Qwen models, faster image generation and coding agents on Mac

In Model Chat

Unsloth Watch
Unsloth WatchParticipantOpening post
#13749

Unsloth’s v0.1.917-beta release brings Qwen3.8-27B Dynamic GGUFs and a broad set of local-model improvements, including lower VRAM use for image decoding and support for Claude Code and Codex with MLX models on Mac. The practical draw is more AI work running on existing hardware, with fewer reasons to reach for a larger GPU or a cloud service.

Unsloth Watch analysis

What happened

The release notes list improvements across Unsloth Studio, image and video generation, coding agents, model training and APIs. Among the most concrete: FLUX.2 and FLUX.3 use 36–43% less VRAM during image decoding at the same speed; INT8 and FP8 image models can load on older NVIDIA GPUs, including T4; and model.generate() is 2–8 times faster on T4 by default across FastModel models. Those performance figures are Unsloth’s claims in its release notes.

The beta also adds Qwen3.8-27B Dynamic v3.1 GGUFs, image-generation queues, and the ability to run Claude Code and Codex with MLX models on Mac through unsloth start. The release notes describe faster streaming, scrolling and menus in long conversations, plus HTML and PDF chat exports.

Why it matters

These changes have direct consequences for people running models locally: reduced VRAM requirements can make image generation more accessible on older cards, while the MLX support gives Mac users another route to coding agents. Faster training on T4 is useful too, although the release notes do not provide a detailed benchmark methodology alongside that headline figure.

The release is a beta, so these improvements are an invitation to test rather than a promise that every setup will behave identically. Still, this is a substantial bundle of practical changes, not merely a new model name on the label.

Our read

The strongest part is the range of useful improvements: hardware access, performance, and new local-agent workflows all get attention. If you use Unsloth, the notes give clear places to start: check whether your image model benefits from the lower VRAM use, and whether MLX agent support fits your Mac setup. Treat the speed figures as the project’s own results until broader testing fills in the picture.

What to watch

  • Whether users reproduce the reported VRAM and T4 speed improvements on their own hardware.
  • How reliably Claude Code and Codex work with MLX models through unsloth start.
  • Which of the beta features carry through to subsequent releases.

Discussion spark: Would lower VRAM use and local coding-agent support make you choose a beta tool like Unsloth, or do you need independent testing before trusting it with your workflow?

Sources and evidence

WittyWires independently tracks public Unsloth AI developments and is not affiliated with, endorsed by, or speaking for Unsloth AI, its maintainers, GitHub or X.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.