Kimi K3.1 API
Kimi K3.1 combines million-token multimodal reasoning with coding, 3D and research workflows, with Pro and faster Highspeed variants.
Kimi K3.1 Pro Highspeed API
moonshot/kimi-k3.1-pro-highspeed
High-speed K3.1 Pro variant with the same advertised intelligence and 1,048,576-token context, approximately 2–3 times faster. Official USD pricing per million tokens: input $4, output $20, cache reads $0.20.
Kimi K3.1 API Background
Overview
Kimi K3.1 is Moonshot AI’s flagship multimodal reasoning family for coding, 3D creation, financial research and long-running agent workflows. It includes Kimi K3.1 Pro and Kimi K3.1 Pro Highspeed.
Development History
K3.1 Pro is a post-trained version of Kimi K3 with expanded reinforcement learning. Both variants are available through Kimi products and the official API. As of October 24, 2026, full model weights are announced for release by November 4, 2026; availability of hosted inference does not mean weights are already downloadable.
Key Innovations
- A 1,048,576-token context window and native image and video understanding.
- Always-on reasoning with low, high and max effort; max is the default.
- Highspeed offers the same advertised intelligence at approximately 2–3 times the speed.
Kimi K3.1 API Technical Specifications
Architecture
A multimodal language model family based on Kimi K3, optimized through additional post-training for reasoning, coding and tool use. Outputs are text and tool calls; 3D assets are produced through code and external tools rather than native 3D API output.
Parameters
Officially described as 3-trillion-class. No exact K3.1 total or active parameter count is asserted here. Both variants support 1,048,576-token context.
Capabilities
- Text, image and video inputs; text responses and tool calling.
- OpenAI-compatible Chat Completions and strict JSON Schema structured output.
- Official model IDs: kimi-k3.1-pro and kimi-k3.1-pro-highspeed.
- Official USD rates per million tokens: Pro input $2, output $10, cached input $0.20; Highspeed input $4, output $20, cached input $0.20. Platform homepage cache-write rates are $2 and $4 respectively; cache write pricing also depends on TTL.
Limitations
- Thinking cannot be disabled. Use reasoning_effort low, high or max.
- Image input uses base64 or uploaded ms:// file references, not ordinary public image URLs.
- Sampling parameters including temperature 1.0 and top_p 0.95 are fixed.
- Output budget shares the context constraint; its configurable maximum is not extra capacity beyond the context window.
- Kimi Code subscription model IDs differ from the pay-as-you-go API model IDs.
Kimi K3.1 API Performance
Strengths
- Moonshot reports 65.7% on Terminal-Bench 4.0 for K3.1 Pro.
- Moonshot reports 61.2% on its internal Kimi Code Bench 3.0 Long-horizon.
- With tools, the vendor reports 92.7% Voxel IoU on its BenchCAD evaluation.
- The vendor reports 64.4 on OfficeQA Pro v2 and 58.6 on Frontier Finance.
Real-world Effectiveness
The family targets repository debugging, multi-step development, parametric CAD scripting and evidence-based financial workflows. Reported benchmark results are vendor evaluations, depend on the harness and budget, and are not guarantees for a particular task. Highspeed is a latency and price variant, not a separately demonstrated intelligence upgrade.
Kimi K3.1 API When to Use
Scenarios
- Use Pro for long-context development and research when token cost matters.
- Use Highspeed for interactive agents where faster responses justify higher input and output rates.
- Use a tool-enabled environment for Blender, CAD, spreadsheets and multi-step workflows.
Best Practices
- Preserve the complete assistant message in multi-turn tool workflows and parse reasoning_content separately from final content.
- Keep stable prompts, tools and documents at the front to improve prefix cache reuse; check cache-write TTL and usage accounting.
- Validate generated code, financial calculations and structured outputs.
- Official references: https://www.kimi.ai/blog/kimi-k3-1-pro ; https://platform.kimi.ai/docs/guide/kimi-k3-1-pro-quickstart ; https://platform.kimi.ai/ .