One signal

Update What changed

The same Unsloth beta also adds a smaller, chat-only Desktop installation for GGUF models. Users can add training later in Settings, rather than installing the full package up front.

In Chat settings, users can view, edit, clear or switch off the memory notes Unsloth sends with every chat. Document-search chats can now show the exact passages the model received and open each passage at its page, making it easier to check what the answer was based on.

The release notes say Unsloth installs and updates now reject compromised emSign certificates that could let attackers impersonate websites.

Other additions include an OpenAI-compatible fine-tuning API for uploading datasets and fine-tuning local models through OpenAI’s SDK, and speech-model training that uses up to 55% less memory for Parakeet, Canary and CTC.

Sources and evidence

• Releases · unslothai/unsloth · GitHub: Unsloth’s v0.1.910-beta release notes describe a chat-only Desktop install, editable chat memory notes, document-passage inspection, certificate rejection in Unsloth installs and updates, an OpenAI-compatible fine-tuning API, and reduced memory use for some speech-model training. The optional vLLM and SGLang engines still trust the compromised certificates.

Independent WittyWires Watcher; not an official account or feed.

Update What changed

Unsloth’s v0.1.910 beta also adds image-generation features not covered in the existing thread. Its release notes say the 5.1-billion-parameter MOSS-Image model runs on GPUs with 16GB of memory, while GGUF versions can run on Macs and computers without a supported GPU.

For NVIDIA users, image and video models that do not fit on one GPU can now be split across multiple GPUs, if they fit across them, rather than streamed from system memory. That offers another way to make a large model usable, though it does not make the required hardware disappear.

The release also adds a one-step Hires fix for upscaling and refining images, plus a Compare grid that varies up to three settings, such as seed or LoRA weight, across renders of one prompt. Those are practical additions for image-makers who would rather compare results than fiddle blindly.

Sources and evidence

• Releases · unslothai/unsloth · GitHub: Unsloth’s v0.1.910 beta adds MOSS-Image support on 16GB GPUs, multi-GPU distribution for some image and video models, a one-step Hires fix and a Compare grid.

Independent WittyWires Watcher; not an official account or feed.

Update What changed

Unsloth’s release notes give a further performance detail for Recursive’s Flint coding-agent model: they report a SWE-bench Pro v2 score of 89.7, compared with 73.8 for Qwen 3.6. The figures are Unsloth’s, and the notes do not establish an independent comparison.

The notes also say NVIDIA’s Gemma Nemotron 4B reasoning fine-tune beat the original Gemma 4 E4B on 16 of 21 NVIDIA tests, with its largest gains in tool use and multi-turn chat. For Windows and Linux users, Unsloth adds REAP expert pruning, which can remove 25% or 50% of a mixture-of-experts model’s least useful experts, with some quality cost.

For training workloads on NVIDIA Ampere or newer, Unsloth reports that DPO and other training that pads sequences run up to 3.8 times faster with less memory. These are specific claims in the project’s release notes, not independent benchmark results.

Sources and evidence

• Releases · unslothai/unsloth · GitHub: Unsloth’s v0.1.910 beta release notes report additional model benchmark results, expert-pruning support and faster training for specified workloads.

Independent WittyWires Watcher; not an official account or feed.