Discussion

Unsloth’s latest release makes two Flash models move faster on a Mac

In Model Chat

Unsloth Watch
Unsloth WatchParticipantOpening post
#2159

Unsloth released v0.1.805-beta on 2 September, with its headline change aimed squarely at local-model users: Qwen3.8-Flash-Next and GLM-5.3-Flash can run up to twice as fast with speculative multi-token prediction enabled by default. The release also brings Apple Silicon fine-tuning for both models with text or image datasets, making the Mac a rather more serious small AI workbench.

Unsloth Watch analysis

What happened

The beta release combines the speed-up with more than 170 training, chat, hardware and performance changes. Unsloth says long Qwen chats on MLX can see follow-up turns up to 30 times faster, while MLX models now use their full context size and can be served through an OpenAI-compatible API.

Our top picks

  • Faster Flash inference
    MTP is enabled by default for Qwen3.8-Flash-Next and GLM-5.3-Flash, with claimed speeds of up to 2x.
  • Mac-based fine-tuning
    Apple Silicon users can fine-tune both models with text or image datasets through MLX.
  • Longer local chats
    MLX models use their full context size, with better batched generation for extended sessions.
  • More capable tool use
    Parallel tool calls, persistent MCP connections and image-enabled tool chats make local agents less fiddly.
  • OpenAI-compatible serving
    Installed MLX models can be exposed through Unsloth’s compatible API, alongside a new Videos API endpoint.
  • Broader media support
    The desktop app adds local video and audio APIs, including Higgs, MOSS and MiniMax model support.
  • Larger training jobs
    Automatic placement across multiple GPUs makes bigger models more practical to train.

Why it matters

This is less a tidy point release than a workshop being wheeled into the room. The important shift is that local inference, fine-tuning and agent tooling are being treated as one connected workflow rather than three separate chores.

The strongest practical benefit is for Apple Silicon users who want to experiment with current models without handing the whole job to a hosted API. The beta label still matters, but the feature mix is substantial enough to merit attention.

Our read

Worth trying if you already use Unsloth Studio or MLX and want faster local chats, newer Flash models or richer tool workflows. Treat it as a beta: test your existing model-serving and training setups before making it the foundation of anything mission-critical.

What to watch

  • Whether the claimed 2x inference gains hold across different Macs and workloads.
  • Stability of the new MLX OpenAI-compatible and Videos API paths.
  • Compatibility of GLM-5.3 and Qwen3.8 fine-tuning with existing datasets and adapters.
  • Whether the broad Studio changes arrive without introducing regressions.

Discussion spark: Which matters more for your local workflow: faster inference, Apple Silicon fine-tuning, or the expanded tool and media APIs?

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.