Moonshot AI Releases Kimi K3.1 Pro, a 3 Trillion Parameter Open-Weight Model

Quick Facts

• Kimi K3.1 Pro has 2.8 trillion total parameters and 104 billion active parameters in a mixture-of-experts architecture, with a 1 million token context window.
• The model scores 55 on the Artificial Analysis Intelligence Index, an 11-point jump over its predecessor K3, ranking fourth behind Claude Opus 5.5, Claude Sonnet 5.5, and Muse Wonder Full.
• API pricing starts at $2.00 per million input tokens and $10.00 per million output tokens, with full model weights scheduled for public release in November.

Moonshot AI launched Kimi K3.1 Pro on October 23, positioning it as the company's most capable open-weight model to date. The Beijing-based AI lab plans to release the full model weights to the public by November.

The model is a post-trained version of Kimi K3, built on the same 2.8 trillion parameter mixture-of-experts base but with substantially more post-training compute applied. Only 104 billion parameters are active per token, using 16 of 896 experts at a time.

Architecture Changes

K3.1 Pro drops traditional positional embeddings entirely. According to Moonshot's technical blog, the model uses NoPE, or No Positional Embeddings, across the full network, replacing the RoPE layers common in competing models. The company says this improves context generalization and computational efficiency.

The model also uses MXFP4 weights with MXFP8 activations through quantization-aware training. The underlying architecture includes Kimi Delta Attention and Attention Residuals, and the model natively processes images and video.

Moonshot is also releasing Kimi K3.1 Pro Highspeed alongside the standard model, which the company says delivers two to three times the inference speed of the standard version.

Benchmark Performance

K3.1 Pro leads the field on several benchmarks. On DeepSWE v1.1, it scores 75.5, ahead of Claude Opus 5.5 at 74.2 and GPT-6 Astra at 74.1. On OfficeQA Pro v2, which tests complex financial document analysis, it scores 64.4 versus 62.2 for Claude Opus 5.5 and 54.4 for GPT-6 Astra.

On Automation Bench-AA, K3.1 Pro scores 73.3, ahead of Claude Opus 5.5 at 69.5 and GPT-6 Astra at 68.5. On HLE-Full with tools, it scores 70.0 compared to 67.7 for Claude Opus 5.5 and 57.2 for GPT-6 Astra.

The model trails on security-related tasks. On ExploitBench, K3.1 Pro scores 74.7 against Claude Opus 5.5 at 91.0 and GPT-6 Astra at 100. Moonshot stripped cybersecurity offense data from its training set, which the company says accounts for the gap.

Hallucination rates dropped sharply from the prior model. On Kimi Hallu Bench Hard, the claim-level hallucination rate fell from 23.4% on K3 to 8.3% on K3.1 Pro.

Coding and 3D Capabilities

On Terminal-Bench 4.0, which tests coding agents on resource-constrained terminal tasks, K3.1 Pro scores 65.7%, ahead of GPT-6 Astra at 58.2% and just behind Claude Opus 5.5 at 66.4%. On Moonshot's internal Kimi Code Bench 3.0 Long-horizon test, it scores 61.2%, up from 55.95% for K3.

The model is built to work inside tools used by game and 3D design studios, including Blender, Unreal Engine, Godot, and CAD software. On BenchCAD, which evaluates 3D reconstruction through code generation, K3.1 Pro scores 92.7% with tools, approaching Claude Opus 5.5 at 96.2% and GPT-6 Astra at 95.9%.

Finance Use Cases

Financial research improved by 11.6 points over K3, and the Business Operation benchmark, which simulates long-horizon management in a business environment, came in at roughly 4.2 times K3's result. Financial modeling improved by 7.2 points.

The model is live now on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. Cached input tokens are priced at $0.20 per million, giving cost-sensitive applications a meaningful discount for repeated context.