AtomicChat releases quantized Ling 3.1 flash models with 86–88% token accuracy
The quantized versions match the base BF16 model’s token predictions 86 to 88% of the time on held-out text, with lower KL divergence scores indicating better alignment. Two 1-bit experimental files are included, but users are advised to test them on their own tasks first.
Key points
- AtomicChat’s quantized Ling 3.1 flash models achieve 86–88% top-1 token accuracy vs. base BF16 model
- Two experimental 1-bit files included, but users must validate performance for their use cases
- Model supports 32,768-token context with 63 linear-attention layers and 9 compressed-attention layers
The builds use an AD-Q4KM quantization scheme, where routed experts are quantized to 1-bit while other layers remain at q8(0) or higher. The model supports 32,768 tokens of context, with 63 layers using linear attention and 9 retaining compressed attention. AtomicChat notes the model’s efficiency: around 0.5 GB per conversation plus 10 KB per token. The project is open-source under the MIT license, with downloads totaling 524 last month.
Model page: Ling 3.1 Flash →
AtomicChat/Ling-3.1-flash-GGUF · Hugging Face
huggingface.co · 15 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 →- Black Forest Labs releases FLUX 3 Image Klein as open-weights model · 2 src
- NVIDIA releases Nemotron 3.5 Super VL and Bolt edge models with open weights · 6 src
- Gemini 4 Argon for Robotics · 2 src
- NVIDIA releases Nemotron 3.5 Super VL 120B with 1M-token context and speculative decoding · 2 src
- Add Nemotron Super 3.5 VL and Nemotron Bolt 74B A9B support by frozenblade1224 · Pull Request #30520 · ggml-org/llama.cpp · 1 src
Comments
via GitHub Discussions