Kimi K3.1 Pro open weights land: self-host a 2.8T near-Opus coder
Moonshot kept its Nov 4 promise: a 2.8T MoE (104B active), 1M context, now on Hugging Face. Ties Opus 5.5 on Terminal-Bench — if you can host ~1.6TB.

Copy markdown
The weights are finally on Hugging Face
Moonshot published Kimi K3.1 Pro's open weights this week — a 2.8T-parameter MoE that activates 104B per token, with a 1M-token context and MXFP4 weights baked in via quant-aware training. It ships under a custom Kimi K3.1 Pro License (not Apache), and providers like Novita are already serving it; API pricing is $2/$10 per million input/output tokens.
Self-hosting it is a ~1.6TB commitment
Day-one vLLM guides are already up, but the hardware math is steep: the MXFP4 weights alone run ~1.6TB on disk, so this is a multi-GPU-server model, not a laptop one. The upside — only 104B of the 2.8T params fire per token, so throughput sits closer to a mid-size model than the headline count implies.
Read the benchmark tables with a pinch of salt
Moonshot's own charts put K3.1 Pro at Terminal-Bench 65.7 (vs Opus 5.5's 66.4) and BrowseComp 92.1 (vs GPT-6 Astra's 91.5) at 'max' reasoning effort. Most are self-reported on internal harnesses; Simon Willison noted it ranked near the top of Artificial Analysis, but that doesn't independently confirm the vendor's tables. Verify on your own workload before you switch coders.