Moonshot's Kimi K3.1 Pro Advances Coding & 3D While Deliberately Curbing Cyberattack Skills

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

Post-Training Scale and RL Infrastructure

K3.1 Pro is not a new architecture but a post-trained version of K3. Moonshot significantly increased post-training compute to support joint training across hundreds of downstream domains. The team optimized every layer of the RL stack — framework, long-context reasoning, training efficiency, sandbox infrastructure, and scheduling — yielding broad gains from large-scale reinforcement learning.

Benchmark Results

On code, agent, and knowledge benchmarks, K3.1 Pro reaches or nears the strongest closed-source models. Specific scores:

Terminal-Bench 4.0 (resource-constrained terminal tasks): 65.7%, ahead of GPT-6 Astra (58.2%) and just below Opus 5.5 (66.4%).

Kimi Code Bench 3.0 Long-horizon (multi-step execution, tool use, end-to-end problem solving): 61.2%, up from K3's 55.95%.

GDPval-AA v2.1 : 1846, surpassing Fable 5.1 (1754) and GPT-6 Astra (1573), trailing Opus 5.5 (1866).

Agents' Last Exam : exceeds both GPT-6 Astra and Opus 5.5.

BenchCAD (3D creation with tools): 92.7%, close to Claude Opus 5.5 (96.2%) and GPT-6 Astra (95.9%).

OfficeQA Pro v2 (financial document analysis): 64.4, current SoTA.

Frontier Finance and internal Financial Modeling, Financial Research, Business Operation benchmarks: all surpass K3, with Business Operation scoring ~4.2× K3.

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

3D Creation and Spatial Reasoning

K3.1 Pro handles Unity, Unreal Engine, Godot, Blender, and CAD scripts, combining web search with 3D generation for complete experiences. The kimi.com and Kimi App now include a "3D Make" feature for online preview, editing, and direct download for 3D printing.

Deliberate Reduction of Cyberattack Capability

A distinguishing focus of the release is safety. Moonshot observed that as post-training scale grew, vulnerability-research ability rose steadily. Root-cause analysis showed cyber-offense content was present in mid-training corpora and RL environments. To mitigate, the team identified and removed cyber-offense data from both mid-training and RL stages while preserving general programming utility.

On the internal xdev-bench-hard benchmark, an unfiltered configuration achieved 90.5% pass@3, whereas the released K3.1 Pro scores 81.0%. Moonshot states its cyber-offense capability still lags behind frontier closed-source models published at least three months earlier.

The company articulated two principles for open-weight releases: (1) do not release weights that advance the absolute frontier of cyber-offense; (2) gradually increase usable cyber capabilities in open weights. The argument: open-weight models can help more organizations audit and harden software, including legacy systems rarely receiving dedicated security reviews, ultimately raising overall software security.

Availability, Pricing, and Early Issues

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/M tokens (cache hit) / ¥13/M tokens (cache miss), output ¥68/M tokens. Highspeed costs 2× standard.

Moonshot plans to release full model weights before November 4, allowing external developers to run and modify the model locally.

Post-launch, some subscribers reported abnormal quota consumption and rate-limiting when running multi-sub-agent tasks, hitting the 5-hour rolling window cap. On October 26, Kimi apologized and announced a one-time reset: all active subscribers receive 25% of monthly quota as compensation (≈ one week). If current usage is below 25%, quota resets to zero. Configuration changes reduce unnecessary sub-agent calls and output length to stretch quota further.

Industry Context

The article concludes that LLM competition is moving beyond raw parameter counts. Vendors are deepening vertical specialization and grounding in real-world deployment, while safety-first controls — such as targeted capability constraints during training — are becoming consensus. Balancing open innovation with risk containment remains the core industry challenge.