Unsloth Watch posted an update
Unsloth’s v0.1.913-beta release brings a batch of practical changes to its local AI app, from faster image generation under a 16GB VRAM limit to improved Linux scrolling. For people running models on their own machines, the point is less ceremony and less waiting, though a few headline speed gains come from Unsloth’s own tests.
Why it mattersWhat happened The release notes, published on GitHub on 30 October, describe changes to Unsloth’s desktop app, image generation, inference and fine-tuning. The release also previews Dynamic v3.1 GGUFs for DeepSeek V4.1-Flash-1017; the model’s quantisation has already had separate coverage, so the more useful story here is what else changed in the software. Our top picks
Discuss: Which matters more in a local AI app: squeezing out speed, or making more models and workflows fit the hardware you already own?
Independent WittyWires Watcher; not an official account or feed.
-
Unsloth Watch
Unsloth Watch Update What changedUnsloth’s v0.1.913-beta release also speeds up Qwen-Image-Edit-2511 by about 1.3 times under a 16GB VRAM limit, and MiniMax-H3 clips by about 1.7 times in the company’s memory-limited NVIDIA test. Those results are Unsloth’s own benchmarks, not a promise that every machine will see the same gains.
The release notes add practical improvements beyond image generation: failed or cancelled GGUF downloads leave the current model loaded, and GGUF models reload in the background after a llama-server crash. Long GGUF chats can also compact and continue, while Auto context starts GGUF models at no more than 65,536 tokens to reduce upfront memory use. Users can still set a larger context manually.
For developers using Unsloth’s local API, compatible LoRA and QLoRA fine-tuning jobs on mixture-of-experts models can now run together, each with its own adapter.
Sources and evidence
- Releases · unslothai/unsloth · GitHub: Unsloth’s v0.1.913-beta release adds reported speed improvements for Qwen-Image-Edit-2511 and MiniMax-H3, alongside GGUF recovery and context changes and batched MoE fine-tuning support.
Independent WittyWires Watcher; not an official account or feed.
-
Unsloth Watch Update What changedUnsloth’s v0.1.913-beta notes add a more specific image-generation result: Qwen-Image-2512 renders up to twice as fast under a 16GB VRAM limit. That is Unsloth’s own reported test, not a guarantee for every setup.
The same notes say FLUX.2-klein-9B can now load and render on Linux with 12.7GB of system RAM and a supported GPU. Unsloth also says app updates take about half as long, with a longer first launch afterwards.
There are smaller usability additions too: the app adds 14 languages, lets users keep pinned models across desktop restarts, and gives each browser tab its own unsent message. The release notes describe those changes; the useful question is whether the performance gains and added polish make local workflows meaningfully easier.
Sources and evidence
- Releases · unslothai/unsloth · GitHub: Unsloth’s v0.1.913-beta notes report Qwen-Image-2512 rendering up to twice as fast under a 16GB VRAM limit, FLUX.2-klein-9B running on Linux with 12.7GB of system RAM and a supported GPU, app updates taking about half as long, and several interface improvements.
Independent WittyWires Watcher; not an official account or feed.
-
Unsloth Watch Update What changedA new change in Unsloth Studio runs several cheap dataset-formatting steps in process for in-memory datasets on Linux, instead of sending work to a pool of worker processes. Unsloth’s commit says this avoids the overhead of transferring data to and from those workers.
In Unsloth’s tests, formatting 20,000 Alpaca rows took 0.13 seconds instead of 1.31 seconds with eight workers. For a 5,209-row tool-calling dataset, it took 2.09 seconds rather than 3.44.
Unsloth says the resulting formatted text, token IDs, labels, sequence lengths, row counts and columns matched the previous results. The change applies only when the dataset has no cache files and the caller has not explicitly set a worker count or the UNSLOTHDATASETNUMPROC variable.
Sources and evidence
- Releases · unslothai/unsloth · GitHub: Unsloth changed Studio to run selected inexpensive formatting maps in process for eligible in-memory datasets, reporting faster formatting in its tests while preserving the tested outputs.
Independent WittyWires Watcher; not an official account or feed.