AtomicChat releases quantized GGUF files for Nanbeige4.3-33B
AtomicChat has released quantized GGUF versions of Nanbeige's 33B dense agent and reasoning model, Nanbeige4.3-33B. Built on release day, 25 October, these files are derived from the original BF16 weights and are available under the Apache-2.0 license. The model features a Llama architecture with a 512K context window and has thinking enabled by default.
Key points
- AtomicChat released quantized GGUF files for Nanbeige4.3-33B on 25 October.
- Custom AD files show significantly lower KL divergence than standard llama.cpp presets.
- The model is sensitive to quantization, with standard Q6K achieving only 95.5% top-1 accuracy.
The release includes both standard llama.cpp presets (Q80, Q6K, Q5KM, Q4KM) and custom "Atomic Dynamic" (AD) files that use a specific importance matrix to preserve quality. According to AtomicChat's measurements, Nanbeige4.3-33B is unusually sensitive to quantization; standard Q6K files only match the original model's next-token predictions 95.5% of the time, whereas the AD-Q6K-Q5K file reduces KL divergence by 39% compared to the standard Q6K.
Users can run these models via llama-server or Atomic Chat. The article notes that files below 4 bits (such as AD-IQ3XXS) suffer significant performance drops, with top-1 accuracy falling below 85%. For optimal results, AtomicChat recommends specific sampling parameters: temperature 0.6 and top-p 0.95 for reasoning and chat, or temperature 1.0 and top-p 0.95 for agentic tasks.
Model page: Nanbeige4.3-33B →
The story so far
2 episodes →- AtomicChat releases quantized GGUF files for Nanbeige4.3-33Bthis story
Nanbeige4.3-33B Quantised GGUF
huggingface.co · 25 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 →- Grok releases first seven minutes of its AI-generated Odyssey series · 1 src
- Daily AI news that isn't swayed by rumors. MetAI 'AI Tsushin' is now live · 1 src
- Moonshot AI releases Kimi K3.1 Pro, jumps to 4th in world ranking · 14 src
- Hugging Face releases Carbon-X, a 14.8B open DNA language model · 1 src
- Am17an merges tensor conversion changes to Qwen in llama.cpp · 1 src
Comments
via GitHub Discussions