Skip to content
AI.info

The Pulse

Unsloth updates Qwen3.8-27B GGUFs with Dynamic V3.1 preview

Unsloth has released updated Qwen3.8-27B GGUF quantizations using its Dynamic V3.1 preview algorithm. The company says most benchmarks show a 5–10% accuracy improvement over its earlier quants at the same size.

Unsloth updates Qwen3.8-27B GGUFs with Dynamic V3.1 preview

Updated Qwen3.8-27B GGUFs are available with Unsloth's Dynamic V3.1 preview quantization algorithm. In its November 11 update, Unsloth said the new quants improve accuracy by 5–10% at the same size on most benchmarks compared with its earlier quants.

Most quantizations from 2-bit through 6-bit have been updated. The 3-bit UD3 versions remain in place: the company found their results comparable to UD3.1 and said the accuracy recovery was not significant enough to warrant replacing them.

What was tested

Qwen3.8-27B is a dense model with 27 billion parameters, vision and reasoning capabilities, and a 256K context window. The updated GGUFs are intended to run locally.

The company's tests included the agentic benchmarks Terminal Bench 2.1 and DeepSWE v1.1, as well as LiveCodeBench v6, GPQA Diamond, MMLU-Pro and AIME25. It used 256K context lengths for the agentic benchmarks and 128K for the others.

For comparisons involving Exllama-v3, the token embedding layer was excluded from the horizontal axis because Exllama-v3 uses BF16 for that layer. The main plots exclude embedding-layer sizes for all quant providers. Unsloth says including those sizes changes the plots when comparing actual RAM and disk use against accuracy.

In a separate comparison with ISTA's GSQ-RCO quants on AIME25, Unsloth says it ran 14 trials and measured average accuracy recovery of 93% to 96% for ISTA's quants. The company estimates ISTA ran one trial and says repeated runs led it to a different conclusion about how the quants compared.

Memory and settings

The company lists 16–19GB VRAM for its 4-bit Qwen3.8-27B quants, citing the RTX 5080 and 4090 as examples; it also lists a Mac with 24GB RAM. Its hardware table gives 16–19GB as the total memory range for 4-bit versions, counting RAM plus VRAM or unified memory.

The same table lists 9–11GB for 2-bit and 12–14GB for 3-bit versions, rising to 23–26GB for 6-bit. Using MTP for faster inference requires 1–2GB of additional headroom, according to the guide.

The model has a maximum context window of 262,144 tokens, which the guide says can be extended to 1M via YaRN. It also supports adjustable reasoning effort and a preserve-thinking setting that retains the previous conversation's thinking trace. The latter uses more tokens but could increase accuracy in continued conversations.

The guide lists xhigh as the default reasoning effort, alongside medium, low and none. Thinking and non-thinking modes also have different default sampling settings.

Running the quants

The updated GGUFs are available through Unsloth's Hugging Face repository, and the model can be run in Unsloth Desktop or llama.cpp. Unsloth Desktop is available for macOS, Windows and Linux; its model hub can be used to search for and download a chosen quant.

The quants also include developer-role support for agentic tools such as Codex, MTP for faster inference and changes to parsing nested objects for tool calling. The guide says users can serve Qwen3.8 through an API using the Unsloth command and llama-server runtime flags.

Sources

Companies in this story

Related stories

Explore

More articles