Moonshot AI Opens Kimi K3.1 Pro Weights to Beat Claude on Coding Agents

Moonshot AI has published the full weights for Kimi K3.1 Pro on Hugging Face, roughly two weeks after launching the model through its API. The release allows teams to inspect, fine-tune, and self-host the model, subject to its license and substantial hardware requirements.

Moonshot’s benchmark results place K3.1 Pro within 0.7 points of Claude Opus 5.5 on Terminal-Bench 4.0 and ahead on several software-engineering evaluations. Those comparisons apply to specific agent harnesses and tasks rather than overall model quality.

Same backbone, heavier post-training

K3.1 Pro retains the Kimi K3 architecture while using substantially more post-training compute. Moonshot says that process included joint reinforcement learning across hundreds of downstream domains, with the goal of improving tool use, planning, and task completion.

• Architecture: Mixture of experts, unchanged from Kimi K3
• Total parameters: Approximately 2.8 trillion
• Active parameters: Approximately 104 billion per token
• Context window: 1,048,576 tokens
• Inputs: Text and images
• Inference mode: Thinking mode only

A mixture-of-experts model routes each token through a subset of its parameters, reducing inference compute relative to a dense model of the same total size. A deployment still needs access to the complete weight set. The raw weights would occupy roughly 5.6 TB in BF16 or 2.8 TB at 8-bit precision, before runtime overhead and the key-value cache required for long contexts.

The API exposes kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Moonshot describes Highspeed as functionally identical and two to three times faster. Both accept thinking-mode requests exclusively; disabling thinking produces an HTTP 400 response.

Moonshot says it removed cyber-related data from both documented training stages while retaining general-purpose coding material. That curation narrows the training set, although deployers still need application-specific security and misuse testing.

At publication, Moonshot had not announced a separate license for K3.1 Pro. Kimi K3 uses a custom, non-OSI-approved license, so teams should verify commercial use, modification, and redistribution terms in the repository before deployment.

Benchmarks favor coding agents

Moonshot’s strongest results come from agentic coding evaluations, where a model works through multi-step tasks using terminals, repositories, and other tools.

Artificial Analysis separately reported a 67.2% Terminal-Bench 4.0 score for K3.1 Pro, the highest result in its cited comparison. Harness differences limit model-only conclusions: K3.1 Pro ran mainly through Claude Code, while other models may have used different scaffolding. The result therefore measures the combined model, agent harness, prompts, and tool configuration.

Science, desktop, and CAD expose the gaps

Moonshot’s results show a narrower capability profile outside software engineering. K3.1 Pro scored 38.1 on Terminal-Bench Science 0.1, compared with 63.3 for Claude Opus 5.5 and 68.1 for GPT-6 Astra.

Additional Moonshot comparisons place K3.1 Pro behind Opus 5.5 on OSWorld V2.1 and BenchCAD, and behind the Claude and GPT entries on ProgramBench and Humanity’s Last Exam without tools. It leads Opus 5.5 on SWE-Marathon and Agents’ Last Exam. The published results favor repository-scale coding over scientific terminal work, desktop control, and CAD generation.

API pricing shifts the cost equation

Moonshot launched the API on October 23, 2026, at $2 per million input tokens and $10 per million output tokens, one-third below Kimi K3’s stated rates. Filling the 1,048,576-token context would cost about $2.10 for input alone. The Highspeed variant costs roughly twice as much per token in exchange for lower latency.

Developers can access the hosted model through the Moonshot API or OpenRouter. Self-hosting replaces per-token charges with cluster capacity, storage, networking, and operations costs, making utilization and data-control requirements central to the decision.

Open weights widen deployment choices

Open-weight access gives teams more control over where code and prompts are processed. Organizations with data-residency rules can run the model on their own infrastructure, while researchers can inspect or fine-tune it for specialized codebases if the final license permits those uses.

Moonshot reports a 63.2-point Terminal-Bench 4.0 improvement over Kimi K3 despite retaining the architecture, parameter count, and context window. If the benchmark version, harness, prompts, and scoring setup are comparable, the gain indicates that reinforcement learning and post-training data changed agent behavior substantially without a new backbone.

Where the model fits

Production checks

1. Confirm the license permits the intended commercial, modification, and redistribution uses.
2. Size the cluster for the chosen precision, expert parallelism, peak context length, and concurrent requests.
3. Enable thinking mode throughout the client stack and handle HTTP 400 responses explicitly.
4. Reproduce representative tasks with the planned agent harness instead of relying on vendor benchmark scores alone.
5. Run security, privacy, and failure-recovery evaluations against production repositories and tools.