Unsloth’s version 0.1.910-beta adds support for Recursive’s Flint, OpenMOSS’s MOSS-Image and NVIDIA’s Nemotron Bolt 4B. The release says users can run the models locally through Unsloth. It also adds a smaller, chat-only Unsloth Desktop installation; users can add training features later in Settings.
Flint is a 40-billion-parameter mixture-of-experts model, a design that activates only some of its parameters for each task. Unsloth says Flint uses 2.5 billion active parameters, supports a 256,000-token context and can fit in 24 GB of memory at 3-bit quantization for short contexts. The release reports a score of 89.7 on SWE-bench Pro v2, compared with 73.8 for Qwen 3.6, but does not give test settings or identify who ran the evaluation.
Image generation and local fine-tuning
MOSS-Image is a 5.1-billion-parameter image model that can generate images containing English or Chinese text. Unsloth says it runs on GPUs with 16 GB of memory; its GGUF versions can also run on Macs and computers without a supported GPU. The release does not provide image-quality comparisons or performance figures.
Nemotron Bolt 4B is NVIDIA’s reasoning-focused tune of Gemma 4 E4B, intended for tool use and instruction following. Unsloth reports a Tau2-Bench score of 63.2, versus 38.7 for the original model, and says Gemma Nemotron 4B beat Gemma 4 E4B on 16 of NVIDIA’s 21 tests. These are release-reported results, and the post does not include test settings or independent validation.
For people who train models, Unsloth also says its new fine-tuning API works with OpenAI’s SDK to upload datasets and run supervised fine-tuning on local models. The release describes Unsloth Desktop as free and open source, with downloads for Windows, macOS and Linux.
Why it matters
For people with a 16 GB GPU, MOSS-Image may offer a way to generate images with English or Chinese text using Unsloth, according to the release. Flint’s 24 GB memory figure applies only at 3-bit quantization and short contexts, and the release does not provide independent test results or enough settings to compare its benchmark scores reliably.
