Unsloth updates Qwen3.8-27B GGUFs with Dynamic V3.1, claims 5-10% gain

Unsloth, which publishes quantized builds of open models, posted an update on its Qwen3.8 how-to-run page dated November 11th 2026. It has released new Qwen3.8-27B GGUFs built with an updated Unsloth Dynamic V3.1 (preview) algorithm. The company says this improves its quants by 5-10% accuracy at the same size on most benchmarks, and calls the result best in class for most tasks, including coding, agentic workloads and general chat. The intro phrases the same claim as an improvement in "average accuracy" at comparable sizes, including on Terminal Bench and DeepSWE. The source does not say whether 5-10% means percentage points or a relative gain.
Qwen3.8-27B is described as a 27B-parameter dense model with vision and reasoning, a 256K context window (262,144 tokens maximum, extendable to 1M via YaRN), and it is said to run locally on 17GB of RAM/VRAM. Unsloth says its 4-bit quants work on 16-19GB of VRAM, such as an RTX 5080 or 4090, or on a Mac with 24GB of RAM. It thanks Qwen for day zero access.
Most quants from 2-bit to 6-bit were updated. The 3-bit UD3 quants were not, because Unsloth found them on par with UD3.1. It says it is still working on making all quants better. For benchmarking, it ran agentic tests including Terminal Bench 2.1 and DeepSWE v1.1, plus general ones such as LiveCodeBench v6, GPQA Diamond, MMLU-Pro and AIME25, among many others. Agentic tests used 256K context and the rest used 128K. It followed the same methodology as for DeepSeek-V4.1-Flash and Qwen3.8-Flash-Next, which also used the V3.1 preview algorithm. The available text gives no per-benchmark scores.
The quants also carry practical changes: developer role support for agentic tools like Codex, MTP enabled for fast inference, and improved parsing of nested objects in tool calls. Unsloth adds that it is working on an even better version of Dynamic quants, to be applied to Qwen4 and future models.
The page also covers NVFP4. Unsloth is releasing dynamic NVFP4 Qwen3.8-27B quants that it says run about 1.5x faster than BF16 checkpoints, with FP8 KV cache calibration for 2x longer context. They need NVIDIA Blackwell GPUs (RTX 50 series, DGX Spark, B200, B300); older GPUs should use the GGUFs. The page says NVFP4 runs in vLLM only for now and that SGLang is not supported, yet it later gives SGLang instructions that require version v0.5.19. It also cites 92% to 97% accuracy recovery against BF16, measured by KLD and Top-1 agreement, but that passage sits next to Qwen3.6 benchmark text, so it is unclear which model the figure covers.
Finally, the page covers Qwen3.8-2.4T-A95B, a 2.4T-parameter model (95B active) that Unsloth says rivals GPT-5.6 Sol. Full precision needs 4.9TB of storage. Unsloth's 1-bit Dynamic GGUF takes 397GB (91% smaller) and the larger IQ1_S takes 508GB. Unsloth says it extended llama.cpp's IQ1_S from 1.5625 to 1.1875 bits per weight by cutting codebook entries. It says this works well for large models and for post-training quantization without QAT or QAD. The 397GB quant needs at least 450GB of RAM, and the page cites about 20 tokens/s generation on B200s if the model fits.
The page also explains how to run Qwen3.8 in Unsloth Desktop, an open-source app, and in llama.cpp, with recommended sampling settings for thinking and non-thinking modes.
Key facts
- Unsloth released new Qwen3.8-27B GGUFs on November 11th 2026 using its Dynamic V3.1 (preview) algorithm, claiming 5-10% better accuracy at the same size on most benchmarks.
- Most 2-bit to 6-bit quants were updated; 3-bit UD3 quants were left alone because they matched UD3.1.
- Benchmarks named include Terminal Bench 2.1, DeepSWE v1.1, LiveCodeBench v6, GPQA Diamond, MMLU-Pro and AIME25, run at 256K context for agentic tests and 128K for the rest.
- 4-bit quants of the 27B dense model are said to fit 16-19GB of VRAM; new NVFP4 quants for Blackwell GPUs are claimed to run about 1.5x faster than BF16.
- The same page covers Qwen3.8-2.4T-A95B, with a 397GB 1-bit GGUF built on new 1.1875 bits-per-weight data types.
Why it matters
Quantization decides whether a large model fits on hardware people own, and how much quality is lost when it does. Unsloth says V3.1 recovers a meaningful amount of that lost accuracy at the same file size, including on benchmarks such as Terminal Bench and DeepSWE. It also says the same algorithm was used earlier for DeepSeek-V4.1-Flash and Qwen3.8-Flash-Next, and that an even better version is planned for Qwen4 and later models. The page also shows 1-bit data types that bring a 2.4T-parameter model down to 397GB.
Who it affects
People running Qwen3.8-27B locally through llama.cpp or Unsloth Desktop, especially on 16-19GB GPUs such as an RTX 5080 or 4090, or on a Mac with 24GB of RAM. Owners of NVIDIA Blackwell cards (RTX 50 series, DGX Spark, B200, B300) can use the NVFP4 builds. Users of agentic tools like Codex get developer role support and better tool-call parsing. The 2.4T section is for anyone with roughly 450GB or more of RAM.
How to use it
The quants are on Hugging Face and ModelScope as Qwen3.8-27B-GGUF and Qwen3.8-27B-NVFP4. Run the GGUFs in llama.cpp or in Unsloth Desktop, an open-source app for macOS, Windows and Linux that picks inference settings automatically. Recommended thinking-mode settings are temperature 1.0, top_p 0.95, top_k 20, min_p 0.0 and presence_penalty 0.0. Instruct mode uses temperature 0.7, top_p 0.80, top_k 20, min_p 0.0 and presence_penalty 1.5. The default reasoning effort is xhigh, and lower settings are available. NVFP4 quants need Blackwell GPUs and run in vLLM; the SGLang route, per the page, requires v0.5.19. The text gives no licence terms or prices.
How solid is it
This is a first-party documentation page. Every accuracy figure is Unsloth's own, and no independent test or third-party benchmark is cited. The available text gives no per-benchmark scores for V3.1 versus the earlier quants, so the 5-10% claim cannot be checked from it. The page describes the algorithm as a preview. It also words the gain two ways, as average accuracy at comparable sizes and as an improvement on most benchmarks at the same size.
Risks and caveats
The source does not say whether 5-10% means percentage points or a relative gain. The 92-97% NVFP4 accuracy recovery is not clearly tied to Qwen3.8-27B, since nearby benchmark text refers to Qwen3.6. The page contradicts itself on SGLang: it says NVFP4 runs in vLLM only and SGLang is not supported, then gives SGLang steps. The hardware tables appear without column headings in the available text, so only the stated ranges (such as 16-19GB for 4-bit) are reliable. The claim that the 2.4T model rivals GPT-5.6 Sol comes from a garbled sentence on the page.
“We're working on an even better version of Dynamic quants which will be applied to Qwen4 and future models.”
— Unsloth, Qwen3.8 how-to-run page