Kimi K3.1 Pro

TL;DR

Moonshot AI's Kimi K3.1 Pro is Kimi K3 (2.8T total, 104B active mixture-of-experts, 1M-token context, multimodal) with much more post-training compute: large joint RL across "hundreds of downstream domains." It scores 55 on the Artificial Analysis Intelligence Index, fourth overall, ahead of Fable 5.1, GPT-6 Astra and Gemini 4 Argon (53) and behind only three Anthropic and Meta models, led by Opus 5.5 (58). On Terminal-Bench 4.0 it scores 65.7% against Opus 5.5's 66.4% and GPT-6 Astra's 58.2%. API price is $2 in / $10 out per million tokens, the same as GPT-6.1 Sol and half of Opus 5.5. A Highspeed variant runs 2-3x faster at twice the price, except cached input which costs the same. Weights are promised by November 4. If they ship, it is the first (soon) open-weight model this close to the frontier.

Same K3 body, much more RL

The architecture is unchanged; the gain is post-training compute plus the infra that makes it stable.

flowchart LR
  K["Kimi K3 base<br/><small>2.8T MoE, 104B active</small>"] --> A["Hybrid attention<br/><small>KDA state + MLA cache</small>"]
  A --> R["Joint RL<br/><small>hundreds of domains</small>"]
  F["Cyber data filter<br/><small>removed from mid-train + RL</small>"] --> R
  R --> P["K3.1 Pro<br/><small>AA index 55, 1M ctx</small>"]
  P --> H["Highspeed tier<br/><small>2-3x faster, 2x price</small>"]
  classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
  classDef core fill:#e5dbff,stroke:#6741d9,color:#1b1b1b,stroke-width:2px
  classDef loop fill:#fff3bf,stroke:#f08c00,color:#1b1b1b,stroke-width:2px
  classDef exit fill:#d3f9d8,stroke:#2f9e44,color:#1b1b1b,stroke-width:2px
  classDef err fill:#ffe3e3,stroke:#e03131,color:#1b1b1b,stroke-width:2px
  class K input
  class A,R core
  class F err
  class P,H exit
  linkStyle 2 stroke:#e03131,stroke-width:2px

Blue is the inherited base, purple is the model and its RL, red is the safety filter, green is what shipped.

Key claims

Post-training, not pretraining. Moonshot says it rebuilt the RL framework, long-context inference, rollout scheduling, sandboxes and an environment-quality platform so each increment of RL compute "delivers stable, predictable returns." This is an RL-infrastructure release.
Benchmarks (vendor + third party): AA Intelligence Index 55 (+11 over K3); Terminal-Bench 4.0 65.7%; Kimi Code Bench 3.0 long-horizon 61.2% (K3: 55.95%); OfficeQA Pro v2 64.4 vs Opus 5.5 62.2; Agents' Last Exam 43.2; HLE with tools 70.0. Weak spot: Terminal-Bench Science 38.1 vs GPT-6 Astra 68.1 and Opus 5.5 63.3.
Hallucination under tools: false claims fall from 23.4% (K3) to 8.3%, between GPT-6 Astra (5.5%) and Opus 5.5 (15.0%), on Moonshot's internal test.
Cyber filtering: exploit skill rose with post-training scale, so Moonshot removed cyber-related data from mid-training and RL. On internal xdev-bench-hard pass@3 fell from 90.5% to 81.0%, still far above K3's 33.3%. A third-party GPT-5.6 Sol run solved 100%.
Pricing: $2 / $10 per M tokens; Highspeed 2-3x faster at 2x price with cached input at parity.

Why the efficiency angle matters

K3's architecture (documented in the wiki's kimi-k3-in-c summary) mixes 69 KDA layers (Kimi Delta Attention, a linear-attention layer with a fixed-size recurrent state) with 24 MLA layers (multi-head latent attention, a compressed KV cache). That is why a 1M-token context is economically servable: most layers keep a constant-size state, and the remaining cache is latent-compressed. 104B active out of 2.8T total (about 3.7%) is what lets Moonshot price at half of Opus. The Highspeed tier pricing cached input at parity is a tell that its speedup comes from decode-side resources (more replicas or speculative decoding), not from cheaper prefill.

Relation to prior wiki

Extends the open-weight frontier race tracked since Muse Spark Open (10-16), where Meta opened a near-frontier model. K3.1 Pro narrows the open-to-closed gap on the AA index to three points.
Confirms the "post-training compute is the lever" thread: like the 10-23 Harness Scaling paper (train RL across many harnesses for transfer), Moonshot credits breadth of RL domains, not model size.
Cyber filtering contrasts with the containment approach in today's containment summary: Moonshot removes the skill from data; Anthropic and OpenAI keep the skill and contain the model.

Open questions