Kimi K3.1 Pro: Benchmarks, API Changes, and Open Weights

Coding and 3D Creation

K3.1 Pro works across game engines such as Unity, Unreal Engine and Godot, combining them with web search and 3D generation. Moonshot says its front-end pages, interactive Three.js scenes and 3D assets show stronger composition, typography and animation, and every coding showcase in the post was built end to end by the model.

In one internal long-horizon coding test, K3.1 Pro built a human walking simulator with training and evaluation pipelines on top of human-atlas, an open-source anatomical model:

• Scope - 167 missions and 420 independent reviews, with the model analyzing results, diagnosing problems and refining the walking controller between runs.
• Result - The simulated human went from collapsing within seconds to walking with alternating strides for 11.2 seconds.
• Debugging - When reward shaping stopped improving arm motion, the model ran open-loop probes of the simulator, found and fixed 36 errors in its shoulder-muscle moment-arm calculations, and then retrained the controller.

For 3D work, K3.1 Pro writes and runs scripts for Blender and CAD software directly, without MCP connectors. Moonshot also launched 3D Make on kimi.ai and the Kimi app, where you preview and edit the models it creates and download them ready for 3D printing.

Note: K3.1 Pro's showcase apps are games, Three.js scenes and front-end pages. TestMu AI's Kane CLI checks a running page like these in real Chrome and seals per-step screenshots, network logs and console output into one evidence pack you can review. Create a free TestMu AI account

Finance and Factual Accuracy

Moonshot says financial analysis and investment research demand a high degree of factual accuracy, and it made targeted improvements there. It reports these results against Kimi K3 on its internal finance and accuracy tests:

• Financial modeling - 57.6 against 50.4 for K3, a 7.2-point gain on formula linkage, financial logic and final workbook quality.
• Financial research - 70.4 against 58.8, an 11.6-point gain on producing auditable research under evidence and cutoff constraints.
• Business operation - About 4.2x K3's score on long-horizon management and decision-making in a simulated business.
• Hallucination rate - On Kimi Hallu Bench Hard, the claim-level hallucination rate fell from 23.4% for K3 to 8.3%. Moonshot notes the prompts are built to provoke errors, so these are not everyday error rates.
• OfficeQA Pro v2 - 64.4, the top score in Moonshot's table, on fiscal analysis over real financial documents supplied only as page images.

Agentic and Knowledge-Work Benchmarks

Outside coding, K3.1 Pro's strongest rows are web research and document-heavy agent tasks. GDPval-AA is an Elo score, so its numbers sit on a different scale from the others.

Benchmarks Where K3.1 Pro Trails

Across all 47 rows in Moonshot's table, K3.1 Pro has the top score on 16. These are the rows to rerun on your own workload before you switch:

• ProgramBench - 31.5 against 65.0 for Claude Opus 5.5. Moonshot reports a binary score, which counts a task only when its ProgramBench score reaches 95%.
• Terminal-Bench-Science 0.1 - 38.1 against 68.1 for GPT-6 Astra and 63.3 for Claude Opus 5.5.
• HLE without tools - 51.1 on HLE-Full and 39.4 on HLE-Diamond, behind Claude Opus 5.5 (59.8 and 60.9) and GPT-6 Astra (53.6 and 65.8).
• PostTrainBench v1.2 - 36.5, below all three proprietary models in the table.
• Kimi Code Bench 3.0 Long-horizon - 61.2 against 66.5 for GPT-6 Astra on Moonshot's own long-horizon coding test.
• Kimi Hallu Bench Hard with tools - 8.3, the second-lowest hallucination rate, behind GPT-6 Astra at 5.5.

K3.1 Pro's agentic coding rows ran in the Claude Code harness at max effort, so a run in Kimi Code or at the low effort setting can land somewhere else. Benchmark the harness and effort level you will ship.

API Behavior That Affects Your Tests

The limits section of the quickstart changes how a test suite should treat the model:

• Thinking is always on - K3.1 Pro cannot skip its reasoning. reasoning_effort takes low, high or max, and max is the default, so pin the level your product ships with in every test.
• Sampling is fixed - temperature 1.0, top_p 0.95, n 1 and zero presence and frequency penalties are fixed, and the docs say to omit them. A test that sets temperature to 0 to get a repeatable answer gets no repeatability here.
• Output budget - max_completion_tokens defaults to 131,072 and can go up to 1,048,576. Set it explicitly so test runs and production use the same budget.
• Full assistant messages - Multi-turn conversations and tool calls must send back the complete assistant message unchanged, not only its content. Add a multi-turn tool scenario that checks your client does this.
• Structured output - json_schema with strict set to true constrains message.content. Parse that field, not reasoning_content.
• Vision input - Public image URLs are not supported. Send images as base64 or as an uploaded file ID, with content as an array of objects rather than a serialized string.
• Web search - Moonshot's official web search tool is being updated and is not recommended for production workflows yet, so an agent that searches needs its own tool for now.

Kimi K2 Retirement and Migration

Kimi's models page sets the deadline for teams still on the K2 series. From October 31, 2026, K2 models are no longer available to newly registered users, and Moonshot plans to retire all K2 series models by the end of December 2026.

Moving from kimi-k2.6 or kimi-k2.7-code changes more than the model name:

• Context window - Both K2 models list a 256K context. K3.1 Pro takes 1M tokens, which changes how much history an agent carries before it compacts.
• Thinking mode - kimi-k2.6 supports a non-thinking mode, and K3.1 Pro has none. Prompts that skipped reasoning on K2.6 now reason first, so retime any test with a latency threshold.
• Effort default - K3.1 Pro runs at max effort unless told otherwise. Set the same effort level on both models before you compare their results.

The gates in LLM regression testing cover which checks to keep when the model under an agent changes.

Verifying AI-Written Code in a Real Browser

I work on the computer-vision side of TestMu AI's testing agents, and a rendered 3D scene is the clearest case of output a unit test cannot see: the build can pass while the canvas stays blank. Kane CLI runs a plain-English objective in a real Chrome browser and returns a pass or fail that a coding agent can read inside its own loop:

npm install -g @testmuai/kane-cli
npx @testmuai/kane-cli-skill
kane-cli run "Open http://localhost:5173, drag the 3D product viewer to rotate it, and verify the product model stays visible and the Add to Cart button is enabled" --agent --headless

The skill installer covers Claude Code, Codex CLI and Gemini CLI. Kimi Code is not on that list, so point it at the Kane CLI AGENTS.md instruction file, as the guide to Kane CLI with AI coding agents recommends for builders outside that list. With --agent, each run streams one JSON event per line and ends with a run_end event that holds the status, a summary and the session paths.

Retesting Chat Agents After the Switch

Moonshot's hallucination figures come from its own adversarial prompts, not from your knowledge base or your users. TestMu AI's Agent Testing generates test scenarios from your agent's own documents and scores every reply on nine quality dimensions, so a chat agent moved from K2.6 to K3.1 Pro gets a before-and-after comparison on the conversations it handles.

• Hallucination Detection - Flags statements the agent's knowledge base or conversation context does not support, the failure Kimi Hallu Bench measures at the model level.
• Context Awareness - Checks that the agent keeps and uses details from earlier turns, which matters more once the agent can carry up to 1M tokens of history.
• Production-readiness verdict - Each run ends in Green, Yellow or Red, so the model switch gets a go or no-go call instead of a folder of transcripts.