Anyone knows anything about this Quantized K3.1m Worth it?
A Reddit user posted about a quantized, expert-pruned checkpoint derived from moonshotai/Kimi-K3.1-Pro, noting it removes 75% of routed experts and is intended for vLLM serving and evaluation smoke tests. The user states it is a calibration-free, safetensors-only prototype with expected large quality regression relative to the source model. The checkpoint loads and serves on eight NVIDIA H200…
Key points
- Checkpoint removes 75% of routed experts, retaining 224 per MoE layer from original 896
- On-disk size is about 443G across 96 safetensors shards plus weight index
- Served on 8x NVIDIA H200 GPUs with TP=8, steady GPU memory ~127,797 MiB per GPU
Model page: Kimi K3.1 Pro →
Anyone knows anything about this Quantized K3.1m Worth it?
huggingface.co · 5 November 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 Research
All →- Researchers link AI agent power to human Goal attainment in formal models · 1 src
- PromptOps framework proposed for iterative LLM judge optimization · 1 src
- Researchers introduce BCSS to scale symbolic models of reasoning · 1 src
- Researchers test Cross-Agent formalization for clinical trial criteria · 1 src
- Researchers introduce efxplorer for Cooperative Multi-Agent reinforcement learning · 1 src
Comments
via GitHub Discussions