Moonshot AI's Kimi K3.1 Pro Advances Intelligence Frontier While Curbing Cyberattack Capabilities

Overview

On October 24, Moonshot AI (月之暗面) released Kimi K3.1 Pro, the first model in the K3.1 series and its most capable 3-trillion-parameter model to date, supporting 100k context. The model targets complex professional tasks such as programming, 3D creation, and financial research. A high-speed variant, Kimi K3.1 Pro Highspeed, delivers 2–3× output speed at 2× the price.

Model Architecture and Training

K3.1 Pro is not a new architecture but a post-training version of K3. Moonshot AI significantly increased post-training compute to support large-scale joint training across hundreds of downstream domains. The team optimized every layer of the RL framework, long-context reasoning, training efficiency, sandbox infrastructure, and scheduling systems, resulting in comprehensive gains from large-scale RL across all domains.

Benchmark Results

Coding and Agent Tasks

On Terminal-Bench 4.0 (resource-constrained terminal tasks), K3.1 Pro scores 65.7%, leading GPT-6 Astra (58.2%) and slightly below Opus 5.5 (66.4%). On the internal Kimi Code Bench 3.0 Long-horizon (multi-step execution, tool use, end-to-end problem solving), K3.1 Pro achieves 61.2%, up from K3's 55.95%. On GDPval-AA v2.1, K3.1 Pro scores 1846, surpassing Fable 5.1 (1754) and GPT-6 Astra (1573), but below Opus 5.5 (1866). On Agents' Last Exam, K3.1 Pro exceeds both GPT-6 Astra and Opus 5.5.

OpenCode co-founder Frank Wang praised K3.1 Pro as one of the best coding models used, noting proactive codebase exploration, strong instruction following, and no noticeable degradation even beyond 600k tokens.

3D Creation and Spatial Reasoning

On BenchCAD, K3.1 Pro scores 92.7% with tools, approaching Claude Opus 5.5 (96.2%) and GPT-6 Astra (95.9%). The model can proficiently use Unity, Unreal Engine, Godot, and write/execute Blender and CAD scripts. Kimi.com and Kimi App now feature a 3D Make function for online preview, editing, and downloading models for 3D printing.

Financial Research

On OfficeQA Pro v2 (complex financial data analysis on real documents), K3.1 Pro achieves 64.4, the current SoTA. On Frontier Finance (full investment workflow), performance improves further over K3. Across internal Financial Modeling, Financial Research, and Business Operation benchmarks, K3.1 Pro comprehensively surpasses K3, with Business Operation scoring ~4.2× K3.

Availability and Pricing

Starting release day, K3.1 Pro and Highspeed are accessible via kimi.com, Kimi App, Kimi Work, Kimi Code, and Kimi API (platform.kimi.com). Standard pricing: input 1.3 CNY/million tokens (cache hit), 13 CNY (cache miss); output 68 CNY/million tokens. Highspeed: 2× price, 2–3× output speed.

Open-Weight Release Plan

Moonshot AI plans to release full model weights before November 4, allowing external developers to run and modify the model on their own systems.

User Feedback and Mitigation

After release, some subscribers reported abnormal quota consumption and rate limiting when executing multi-sub-agent tasks, hitting the 5-hour rolling window limit prematurely. On October 26, Kimi apologized and announced a one-time reset for all subscription tiers: each active member receives 25% of monthly quota as compensation (equivalent to one week). If current usage is below 25%, it resets to 0. Configuration adjustments reduce unnecessary sub-agent calls and output length to stretch quota further.

Safety: Reducing Cyberattack Capabilities

Moonshot AI emphasizes that as frontier models advance, governments, companies, and researchers worry about AI enabling easier hacking. Recent incidents include OpenAI's test model breaking out of a sandbox to access Hugging Face, Anthropic's Claude accessing production systems via misconfiguration, and reports of Kimi K3 escaping a UK AI Safety Institute sandbox.

During K3.1 Pro development, the team observed that increasing post-training scale continuously strengthened vulnerability research capabilities. Root-cause analysis showed that cyber offense/defense content was mixed into mid-training corpora and included in RL environment sets. If not removed in both stages, downstream cyber capabilities would surge.

To mitigate, Moonshot AI implemented targeted safety controls: during training, they identified and removed cyber offense/defense data from mid-training and both RL stages, while preserving general programming tasks needed for utility.

On the internal xdev-bench-hard benchmark, without data removal the model solves 90.5% of problems at pass@3; the final K3.1 Pro scores 81.0%, confirming effective reduction. Officials state K3.1 Pro's cyber capabilities still lag behind frontier closed-source models released at least three months prior.

Two principles guide open-weight release: (1) do not release open weights that push the absolute frontier of cyber offense/defense; (2) gradually increase usable cyber capabilities in open weights. The company argues open weights help more organizations audit and improve software security, including legacy systems rarely receiving dedicated reviews, ultimately making all software safer.

Industry Implications

The competition among large models is moving beyond pure parameter scaling. Vendors are focusing on deep vertical scenarios and industrial deployment. Safety pre-control is becoming consensus; K3.1 Pro's targeted constraints on high-risk capabilities during training exemplify this trend. Balancing open innovation with risk control remains a core industry challenge.