Kimi K3.1 Pro

Not a new base model: "Kimi K3.1 Pro is a post-trained version of Kimi K3, with substantially larger post-training compute" — how much larger is not quantified. The architecture mostly carries over from Kimi K3, the 2.8T Stable LatentMoE with 104B active and a 1M-token context (93 layers, 896 experts with 16 routed and 2 shared). There is one change, disclosed only in the open-weight model card: the 24 full-attention layers that were Gated MLA in K3 are now Gated DeepSeek Sparse Attention, with a 32-head indexer selecting the top 2,048 tokens, alongside the unchanged 69 KDA layers. What moved is the score: 55 on AA Intelligence Index v4.3.2 against K3's 44 on the same version, bought entirely in post-training. On v5.0 (October 27) it reads 46, which AA reports as the top open-weights score even though the weights are not yet released. Moonshot's own framing is breadth rather than a single win — "consistent gains in every domain", with coding, 3D creation, financial research and agentic use called out.

The weights shipped on time. The launch post promised "Full model weights will be released by Nov 4, 2026", and at filing (Oct 24) nothing was downloadable, so the entry was carried as closed. moonshotai/Kimi-K3.1-Pro appeared on HuggingFace on 2026-11-02, public and ungated. It holds a single checkpoint of about 2.78T parameters, with MXFP4 weights and MXFP8 activations from quantization-aware training, the same precision Moonshot serves, and it has vision processors. It ships with day-0 vLLM, SGLang and TokenSpeed recipes. The Kimi K3.1 Pro License is MIT-style with two conditions. A Model-as-a-Service operator with more than $20M in trailing-12-month revenue needs a separate agreement with Moonshot for commercial use. A product with more than 100M MAU or more than $20M in monthly revenue must display "Kimi K3.1-Pro". It now tops the open-weights intelligence column at AA v5.0's 46.

The cyber-capability ablation is the most unusual thing in the post, because Moonshot measured its own safety intervention and published the counterfactual. "For the open-weight release of K3.1 Pro, we identified and removed cyber-related data from both training stages, while retaining the general-purpose coding tasks needed for model usefulness", defended on access grounds: "We believe making K3.1 Pro available in a non-discriminatory way benefits people who need dual-use capabilities for defensive purposes." On internal xdev-bench-hard (pass@3) the ladder runs GPT-5.6 Sol 100 (third-party evaluation), the same pipeline without cyber risk-reduction 90.5, DeepSeek-V4.1-Flash 85.7, shipped K3.1 Pro 81.0, GLM-5.3 57.1, and K3 33.3. Moonshot reads it as "data removal meaningfully reduces cyber capability relative to what the same training pipeline would otherwise produce", which is true and worth 9.5 points. The other true reading sits in the same column: the shipped, risk-reduced model still scores 81.0 against its predecessor's 33.3, so the generation more than doubled the capability that the removal then trimmed.

On the capability tables it approaches the closed frontier without passing it. Terminal-Bench 4.0: 65.7% against Claude Opus 5.5's 66.4%. On BenchCAD (1,000-file subset, Voxel IoU) with tools it takes 92.7 behind Opus 5.5 at 96.2 and GPT-6 Astra at 95.9 — but without tools it reads 72.8 against Opus 5.5's 73.0, which is the sharper of the two numbers and the one the scaffold is not doing the work in. It leads on finance, 64.4 on OfficeQA Pro v2 (state of the art) and 58.6 on Frontier Finance, and reports an 8.3% claim-level hallucination rate on Kimi Hallu Bench Hard.

Served on Kimi.com, Kimi Work, Kimi Code and the Kimi API at $2.00 / $10.00 per 1M tokens with cache reads at $0.20, a 1M context and up to 943,718 completion tokens. Reasoning is on by default at max effort, with max/high/low/none exposed, and the modality is text, image and video in. A second variant, Kimi K3.1 Pro Highspeed, "delivers 2-3× the speed of K3.1 Pro for more responsive interactions" through the same channels with no separate price published. AA measures 77.1 output tokens/s, a 2.89s time to first token and $2.11 per Intelligence Index task. No technical report and no GitHub repository at filing — K3 shipped one of each. As of 2026-11-05 the open-weight release added a model card but still no technical report.