The Kaitchup releases statistically-lossless Qwen3.8-27B GGUF models
The Kaitchup published three quantized versions of the Qwen3.8‑27B large language model using a customized Statistically‑Lossless Quantization (SLQ) method and packaged them as GGUF files. The three releases target Expected Acceptance Rates (EAR) of 98 %, 98.5 % and 99 % measured on the BF16 reference’s top‑10 tokens, and differ in precision choices such as Q4\K/Q8\0 for EAR 98 and Q6\K/BF16 for…
Key points
- The Kaitchup released three Qwen3.8‑27B GGUF models quantized with SLQ, targeting EAR 98 %, 98.5 % and 99 %.
- Models use Q4/K, Q5/K, Q6/K precision combos and retain 351,008,768 bytes of quantized MTP tensors; no vision tensors included.
- Validation on an NVIDIA GH200 with llama.cpp confirmed text parity; downstream task equivalence not yet established; 239 downloads last month.
Validation was performed with a llama.cpp build on an NVIDIA GH200 GPU (commit 45451c366d4256f94fbc8845645e31a6b9155054), confirming native text scoring and GGUF parity, though image/video inference and speculative decoding were not covered. The releases have not been benchmarked for downstream task performance, and statistical equivalence remains unproven. The models can be downloaded via the Hugging Face CLI, and a simple OpenAI‑compatible server command is provided. The files saw 239 downloads in the last month, and the author invites users to support the work.
The story so far
3 episodes →- The Kaitchup releases statistically-lossless Qwen3.8-27B GGUF modelsthis story
Statistically Lossless Qwen3.8 27B GGUFs
huggingface.co · 27 October 2026
Loading the full article…
This text was published by huggingface.co. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
1source- Reddit discussionreddit.com
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Generative AI & Models
All →- Liquid AI releases six 2B-parameter open-weight models in LFM3 family · 3 src
- Anthropic reports 3.6x jump in token volume after Claude Opus 5.5 launch · 5 src
- Moonshot AI announces Kimi K3.1 Pro with 1M-token context and 3D model creation · 2 src
- OpenAI’s GPT-6.1 Terra scores 1,671 points on Code Arena WebDev leaderboard · 1 src
- OpenAI and Anthropic cut AI model prices by up to 80% amid rising competition from open-weight models · 6 src
Comments
via GitHub Discussions