Unsloth released new lightweight language model files and improved response generation and image model execution on older GPUs. The company says developers using existing NVIDIA GPUs get more image model options and faster responses with default settings.

1. Run Qwen3.8-27B GGUF locally

Released on November 11, this version provides Qwen3.8-27B in GGUF, a compressed model file format. It uses the Dynamic v3.1 distribution format, which Unsloth says can run locally.

Dequantization, which unpacks GGUF values for computation, was also revised. Reconstructed values now match llama.cpp bit for bit and are rounded only once to the compute dtype, the numeric format used for calculations.

Studio's image generation path also gained broader support for loading low-bit GGUF files, which store values using fewer bits. Beyond the new language model files, the release updates how existing image model files are read and processed.

2. Model generation on T4 is 2–8x faster by default

The FastModel family now uses CUDA graphs by default during response generation. These bundle repeated computation flows in advance for reuse. The release notes report 2–8x faster generation on T4, with output tokens unchanged from the previous execution path.

For developers, the key is that the generation execution path changed rather than the amount of output being reduced. The improvement is part of the default path, with no separate acceleration setting to enable.

For the Qwen3.5 and Qwen3.6 families, this accelerated path also supports caches that reuse previous computations.

Gemma 4 generation is also supported.

An optional feature was added to expand cache space during generation as needed. Alongside default generation acceleration, this gives users a choice in how the cache is managed.

3. Image model execution and memory on older GPUs

INT8 and FP8 image models can now be loaded on older NVIDIA GPUs. This broadens support for precision combinations that previously refused to run.

The black image output issue in FLUX.3 was also fixed. According to the company, FLUX.2 and FLUX.3 use 36–43% less GPU memory during image decoding without changing speed.

For FLUX.3 on T4, attention, which computes the importance of input information, now runs in 16-bit floating point. The company reports that this reduced generation time per image by 15% for GGUF and 34% for INT8.

A separate change rescales intermediate values to maintain low precision. The company reports reductions in generation time per image on the same card of 24% for INT8 and 73% for GGUF Q8_0.

Another change points to a system memory copy of the weights, the numeric values learned and stored by the model, instead of copying them again at each step. The company reports 11% less time per image for FLUX.3 GGUF Q8_0 on that card.

Qwen-Image-2.1 now loads the provided INT8 version of its text encoder, which converts text into information for image generation. The company reports that time per image on T4 Turbo with INT8 fell from 82.7 seconds to 31.9 seconds.

These results show reductions in time before and after each change. They cannot form a single performance ranking: models, quantization methods that represent numbers with fewer bits, and modifications differ.

4. Training expands model and video data support

Context parallel training, which splits long contexts across multiple devices, now supports more models. Kimi-Linear and DeepSeek were added.

MiniMax and Gemma 4 text models are also supported.

Training for video-capable vision-language models, which process images and text together, no longer drops the dataset's video column. For developers preparing video training, the key change is correct use of training data rather than speed.

5. Memory budgets adjusted alongside broader compatibility

Alongside more models running on older cards, accounting for memory needed during computation has changed. FP8 compatibility is achieved by using it only for weight storage. FLUX.3 also has a path that keeps stored weights but computes each layer in 32-bit floating point.

On this path, the activation memory budget for intermediate values was doubled based on a measurement table. The release says this keeps the GGUF Q4 denoiser, which reduces image noise, on the T4. Execution planning now accounts for intermediate values as well as weight storage.

The release centers on new GGUF files, faster generation, and more image model options on older cards such as the T4.