Kimi K3.1 Pro — Advancing Open Intelligence via Large-Scale Multi-Domain RL

TL;DR

Kimi K3.1 Pro is a post-trained version of the 2.8T Kimi K3 (Kimi K3 — Open Frontier Intelligence (2.8T MoE with KDA + AttnRes)) that applies substantially larger post-training compute and large-scale joint RL across hundreds of downstream domains, with every stack layer (RL framework, long-context inference, sandbox infra, rollout scheduling, env-quality platform) co-optimized so each RL-compute increment delivers predictable returns. It is positioned as the first in a new K3.1 series, scores 65.7% on Terminal-Bench 4.0 (ahead of GPT-6 Astra 58.2%, narrowly behind Claude Opus 5.5 66.4%) and advances significantly on 3D/CAD, financial, and long-horizon coding workloads. A Highspeed variant delivers 2–3× the throughput; API is live today and full open weights are promised by Nov 4, 2026.

Key claims

Post-training is scaled with large-scale joint RL across hundreds of downstream domains, and consistent gains are reported in every domain [§Advancing the Frontier of Open Intelligence].
On Terminal-Bench 4.0 (realistic terminal coding agent tasks with strict task-specific tests, Claude-Code harness), K3.1 Pro scores 65.7%, ahead of GPT-6 Astra (58.2%) and behind Claude Opus 5.5 (66.4%) [§Advancing the Frontier of Open Intelligence, §Footnotes].
On the internal Kimi Code Bench 3.0 Long-horizon, K3.1 Pro scores 61.2% vs 55.95% for K3 [§Advancing the Frontier of Open Intelligence].
A long-horizon RL case study on the human-atlas simulator shows the model running 167 missions / 420 reviews, taking a walking-controller from <1s stability to 11.2s of continuous alternating-stride walking, including open-loop probing that corrected 36 shoulder-muscle moment-arm errors when reward shaping stalled [§Coding].
On BenchCAD (3D reconstruction from multi-view renderings via CAD code), K3.1 Pro reaches 92.7% with tools, approaching Claude Opus 5.5 (96.2%) and GPT-6 Astra (95.9%) [§3D Modeling and Design].
On OfficeQA Pro v2 (image-rendered PDFs; agents cannot read raw text), K3.1 Pro reports a SOTA 64.4, with “Financial Research” +11.6 and simulated “Business Operation” ~4.2× over K3 on internal benches [§Finance].
Kimi Hallu Bench Hard: claim-level hallucination rate drops from 23.4% (K3) to 8.3%; share of responses with ≥1 factual error is 42.4% of K3’s rate (stress-test benchmark, not representative of daily use) [§Finance].
Cyber-related data was deliberately identified and removed at both mid-training and the RL env mix; on xdev-bench-hard the unmitigated checkpoint solves 90.5% at pass@3 vs 81.0% for shipping K3.1 Pro, framed as a trained-in capability limit rather than a post-hoc filter [§Cybersecurity].
K3.1 Pro Highspeed delivers 2–3× the speed of K3.1 Pro for latency-sensitive interactions; full open weights for both are promised by Nov 4, 2026 [§Introducing Kimi K3.1 Pro, §Availability].
Reported defaults for all benchmark results are reasoning_effort=max, temperature=1.0, top_p=0.95 [§Footnotes].

Method

K3.1 Pro keeps the K3 pretrained backbone — 2.8T MoE with Kimi Delta Attention + Attention Residuals + Stable LatentMoE + Quantile Balancing + Per-Head Muon, MXFP4 weights / MXFP8 activations under QAT, native multimodal, 1M context — and layers much larger post-training compute on top. The headline methodology is large-scale joint RL across hundreds of downstream domains: a shared platform does env-quality checks across all domains, the RL framework / long-context inference / sandbox / rollout scheduling are co-optimized so adding RL compute gives stable returns, and the resulting model gains in every domain simultaneously rather than trading them off. For safety, cyber-related content is actively removed from both the mid-training corpus and the RL environment mix (while general coding is retained), using an internal pipeline Moonshot plans to continue applying “at every training stage.” Evaluation is run under the Claude-Code harness (sometimes Kimi Code or Codex for competitors), with task-specific grading or Gemini-3.5-Flash / Kimi K3 as judges where applicable.

Results

Terminal-Bench 4.0: 65.7% (ahead of GPT-6 Astra 58.2%, behind Claude Opus 5.5 66.4%) [§Advancing the Frontier of Open Intelligence, §Footnotes].
Kimi Code Bench 3.0 Long-horizon: 61.2% vs K3’s 55.95% [§Advancing the Frontier of Open Intelligence].
BenchCAD with tools: 92.7% vs Claude Opus 5.5 96.2% and GPT-6 Astra 95.9% [§3D Modeling and Design].
OfficeQA Pro v2: 64.4 (SOTA on this bench per Moonshot) [§Finance].
Internal finance deltas vs K3: Financial Modeling +7.2, Financial Research +11.6, Business Operation ≈4.2× [§Finance].
Kimi Hallu Bench Hard: claim-level hallucination 23.4% → 8.3%; any-error-in-response rate down to 42.4% of K3’s [§Finance].
xdev-bench-hard (vulnerability discovery + exploit under mitigations, pass@3): 81.0% final vs 90.5% unmitigated pipeline — reported as evidence that mid-training/RL data removal materially reduces offensive-cyber capability [§Cybersecurity].

Why it’s interesting

K3.1 Pro is the first filed public data point on how far pure post-training RL can push an already-frontier open-weights MoE when the full stack (sandbox, rollout, long-context inference, env quality) is co-designed for multi-domain RL — complementing Kimi K3 — Open Frontier Intelligence (2.8T MoE with KDA + AttnRes) on the pretraining side with a Reasoning-RL + Agentic-SWE post-training story. It also sharpens the open-vs-closed frontier on long-horizon coding: Terminal-Bench 4.0 65.7% puts an open-weights model within ~0.7 points of Claude Opus 5.5 on a strict, task-specific coding benchmark, which is a tighter gap than reported for Introducing Muse Spark Open — The Most Capable US Open-Weight Model to Date at the time of filing. The explicit “remove cyber data at mid-training and in the RL env mix, measure the capability delta, publish both numbers” protocol is also notable as a replicable pattern for open-weight safety claims — a direction Open Weights and American AI Leadership‘s policy argument has not yet seen matched by concrete lab practice at this scale.