Moonshot opens Kimi K3.1 Pro weights on Hugging Face (2.8T/104B MoE, 1M ctx, MXFP4); fulfills Nov 4 promise

Why it matters

【Key points】

Moonshot / Kimi via @Kimi_Moonshot at UTC 2026-11-04T16:00 (CST Nov 5 ~00:00) released Kimi K3.1 Pro full weights—fulfilling e/1052’s “by Nov 4” promise. Primary HF: moonshotai/Kimi-K3.1-Pro (~96 safetensors shards).

• Card: open-weight multimodal MoE flagship; ~2.8T / 104B active; ~1M context; Gated DSA vs Kimi K3; MoonViT-V2 (~401M); QAT MXFP4 / MXFP8 activations.
• Engines: vLLM / SGLang / TokenSpeed; API still platform.kimi.ai (`kimi-k3.1-pro`).
• License: Kimi K3.1 Pro License (not Apache/MIT)—keep notice
• MaaS with >~$20M revenue over any consecutive 12 months needs a separate Moonshot deal
• large commercial products need UI attribution.
• Same day: vLLM Day-0; Together is partner amplify.

Vendor table (max): Terminal-Bench 4.0 65.7, DeepSWE v1.1 75.5, HLE-Full w/ tools 70.0, GPQA-Diamond 94.2 vs Fable 5.1 / GPT-6 Astra / Opus 5.5 / DeepSeek-V4.1-Flash-1017 / GLM-5.3 (vendor/cited, need independent checks).

【Our read】

1. e/1052 shipped API/product with weights due Nov 4; today’s HF drop completes self-host—same pattern as MiniMax M3.1 Flash. Substantive China open-weight delivery, not another param dump.
2. 2.8T/104B + 1M multimodal peers DeepSeek V4.1 Flash, GLM-5.3, prior Kimi K3. Vendor coding/agent rows near closed peers on some scores, but harness footnotes mean wait for independent runs. MaaS>$20M + attribution = open weights with commercial guardrails, not permissive OSS.
3. Day-0 vLLM shortens HF→runnable for private long-context multimodal agents; API vs self-host will split. Watch third-party benches, enterprise license clarity, distill/quant ecosystem.
4. Limits: vendor benches; 2.8T MXFP4 self-host is expensive. Do not merge with e/1052 product launch—this is the weights fulfillment only.

【Verifiable】

• huggingface.co/moonshotai/Kimi-K3.1-Pro
• LICENSE on the same repo
• x.com/Kimi_Moonshot/status/2118010528123773297
• e/1052
• vLLM Day-0