Unsloth says Dynamic 3.0 quants beat rivals by 10% at the same size

Unsloth released Dynamic 3.0, an update to its post-training quantization method for Qwen3.8-27B, claiming more than 10% better top-1 accuracy at the same disk size than competing GGUF files, based on its own benchmarks. The Han brothers' method targets local AI users by preserving model behavior at sizes suited to consumer hardware, and includes a new Divergence-300 @32 test to measure output divergence from the BF16 reference.

Unsloth says Dynamic 3.0 quants beat rivals by 10% at the same size Daniel and Michael Han's post-training quants target local AI users, with the headline gains measured through Unsloth's own tests. By RuntimeWire Staff /author/runtimewire-staff · Published Primary source: Unsloth https://unsloth.ai/docs/basics/dynamic-3.0-ggufs Why it matters The Han brothers are turning quantization into a distribution business: every quality gain at a fixed file size expands the hardware that can run open models locally. Daniel Han and Michael Han are releasing Unsloth Dynamic 3.0 https://unsloth.ai/docs/basics/dynamic-3.0-ggufs?ref=runtimewire , an update to the brothers' compression method for Qwen3.8-27B /models/qwen/qwen3.8-27b with GGUF files designed to preserve more of the original model's behavior at sizes suited to local hardware. The release extends a technical bet the Hans have pursued since founding Unsloth https://unsloth.ai/?ref=runtimewire in 2023: model optimization can serve as a distribution layer for open AI. Each reduction in memory and storage requirements puts a model within reach of another tier of consumer machines, giving developers an alternative to paying an inference provider whenever they want to run it. Daniel Han https://www.linkedin.com/in/danielhanchen?ref=runtimewire previously worked at Nvidia and says he made the t-SNE algorithm 2,000 times faster and fixed more than 20 bugs across open models including Llama, Gemma, Mistral and Phi. Michael Han https://www.linkedin.com/in/michaelhan3?ref=runtimewire handles product, design and engineering. The brothers took Unsloth through Y Combinator's Summer 2024 batch https://www.ycombinator.com/companies/unsloth-ai?ref=runtimewire , and YC currently lists the San Francisco operation at eight people. Lightspeed https://lsvp.com/company/unsloth/?ref=runtimewire identifies Daniel as CEO and Michael as co-founder. A larger claim than a smaller file Unsloth says Dynamic 3.0's Qwen3.8-27B quants deliver more than 10% better top-1 accuracy at the same disk size than competing files. That figure comes from Unsloth's own benchmarks and has not been independently reproduced. The underlying method is post-training quantization. Unsloth says it neither retrained Qwen3.8 on its calibration data nor used quantization-aware training or quantization-aware distillation. Instead, Dynamic 3.0 applies a higher-quality importance matrix, revised layer selection and additional quantization techniques after the base model has already been trained. Unsloth assembled the calibration material around coding agents, chat and multilingual tasks. That choice matters because the data used to decide which weights require greater precision can shape where a compressed model holds up and where it deteriorates. A quant tuned around coding prompts may behave differently on other workloads, even when an aggregate accuracy figure looks strong. Unsloth also acknowledges that its headline top-1 measurement has limits. It records whether a quantized model selects the same highest-probability next token as the BF16 reference, leaving the rest of the generated trajectory unmeasured. Dynamic 3.0 therefore adds a company-designed test called Divergence-300 @32. Unsloth used 300 prompts drawn from Terminal-Bench 2.1, DeepSWE, Harbor, MathArena 2025-26 and a collection of non-Latin and long-document tasks. It then compared 32 tokens of greedy decoding from each quant against the BF16 model. The release also reports KL divergence, which measures how far the quantized model's output distribution moves from the higher-precision reference. Those additions make the evaluation more useful than a single next-token score. The broad performance claim still rests on prompts, calibration choices and comparison settings selected by Unsloth. Developers deciding whether to replace an existing quant will need to test the files against their own applications, especially outside coding, chat and the languages represented in the calibration set. Unsloth has published its importance matrix for community testing and variations. That gives outside developers a path to inspect the method, create fine-tunes and run evaluations beyond Unsloth's benchmark suite. The hardware trade The most aggressive file is the 6.2GB UD-IQ1 S one-bit quant. Unsloth says it is 89% smaller than the reference and retains about 72% top-1 accuracy. That reduction puts a 27-billion-parameter model within the storage and memory range of machines that could not approach the full-precision version, with a substantial quality trade attached. The 9.83GB UD-Q2 K XL offers a less severe compromise. Unsloth reports that it beats the next-best comparison by about 8% on top-1 accuracy and can produce a working HTML program, albeit with a JavaScript bug in the example shown in the documentation. For quants below UD-Q2 K XL, Unsloth removed the model's multi-token prediction module to save roughly 500MB. Users who want that capability can load a separate Q4 0 MTP module. The decision turns storage into an explicit option rather than packaging the speculative-decoding component into every smaller file. The Qwen3.8-27B GGUF files https://huggingface.co/unsloth/Qwen3.8-27B-GGUF?ref=runtimewire work with engines including llama.cpp and Unsloth Desktop https://unsloth.ai/docs/desktop?ref=runtimewire . Unsloth says the Qwen3.8 repositories recorded 5.1 million downloads in five days. That metric represents file requests rather than unique users: Hugging Face's counting rules https://huggingface.co/docs/hub/en/models-download-stats?ref=runtimewire count GGUF file requests and can count multiple files pulled during a repository clone. The demand still explains the timing. Open-model releases now produce a second race among toolmakers to package weights for the machines developers already own. The Hans are positioning Unsloth between model labs and local runtimes, where file format, quantization quality and day-one availability determine which weights get tested and adopted. RuntimeWire reported in June /article/unsloth-glm-5-2-local-gguf-studio that Unsloth had applied the same distribution strategy to Z.ai's GLM-5.2 /models/z-ai/glm-5.2 , using GGUF compression and Unsloth Desktop to make a frontier-scale open model more accessible outside data centers. Dynamic 3.0 turns that recurring release work into a named methodology that Unsloth can apply across model families. That approach gives the brothers a useful place in the open-model supply chain. Labs can publish increasingly capable weights, while hardware constraints still decide who can run them. Unsloth's credibility will depend on how well Dynamic 3.0's gains carry across models and workloads once developers begin measuring the files outside Unsloth's test suite.