Updated November 10, 2026
What is Kimi K3.1 Pro?
Kimi K3.1 Pro is Moonshot AI's open-weight multimodal model for coding, long-horizon agents, 3D creation and financial research. It is post-trained from Kimi K3 and starts the K3.1 series. The official API IDs are kimi-k3.1-pro and kimi-k3.1-pro-highspeed. The model card lists 2.8 trillion total parameters and 104 billion activated parameters in a mixture-of-experts architecture, with a 1,048,576-token context window and text-and-image modalities. Weights are available from Moonshot's official repository. Open-weight is the accurate description: its custom license includes commercial conditions that should not be described as unrestricted open-source terms.
Shortlist K3.1 Pro for complex coding and multimodal agent work when hosted access or a serious self-hosting evaluation fits your requirements. Compare the standard and Highspeed variants on actual task cost, latency and the amount of expert correction required.
Official model and release documentation
Specification | Verified detail |
|---|---|
Provider | Moonshot AI |
Standard API ID | kimi-k3.1-pro |
Highspeed API ID | kimi-k3.1-pro-highspeed |
Architecture | Mixture of Experts |
Total parameters | 2.8T |
Activated parameters | 104B |
Context window | 1,048,576 tokens |
Modalities | Text and image |
Weights | Available in official Moonshot repository |
License | Custom Kimi K3.1-Pro license; conditions apply |
Hosted access | Kimi products and API |
What changed and how to approach migration
K3.1 Pro is a post-trained successor to Kimi K3, with larger post-training investment and an emphasis on sustained agentic work. The model card describes an attention change from full-attention Gated MLA layers to Gated DeepSeek Sparse Attention layers for long-context efficiency. Moonshot also offers Highspeed, which it says delivers the same intelligence at roughly two to three times the speed. That is a provider claim, not a guarantee for every prompt length or region. Keep the predecessor guide intact as a separate version reference rather than quietly replacing every mention of K3 with K3.1.
Area | What changed |
|---|---|
Predecessor | Post-trained from Kimi K3 |
Post-training emphasis | Larger joint reinforcement-learning effort |
Attention update | Gated DSA replaces full-attention Gated MLA layers |
Highspeed variant | Provider claims same intelligence at 2-3 times speed |
Migration implication | Test exact IDs, reasoning settings and tool behavior |
Capabilities and practical use cases
The most distinctive uses combine software work with visual or spatial material. A K3.1 Pro task might inspect a codebase and reference image, generate a 3D scene or CAD script, then revise it after a measurable check. Moonshot also emphasizes financial research and long-running engineering. The release's demos use external tools and asset libraries, so the model should not be described as generating every artifact without a surrounding system. For ordinary coding, retain normal repository practices: inspect changes, run relevant tests and evaluate maintainability. A large context window can help carry the repository, but it does not replace selecting the files and requirements that matter.
Workflow | Practical fit |
|---|---|
Long-horizon coding | Carry changes through diagnosis, implementation and review |
Visual understanding | Use screenshots, diagrams and reference images |
3D and CAD workflows | Generate scripts and revise tool-produced assets |
Financial research | Analyze sources and maintain numeric consistency |
Agent tools | Use documented tool-calling and reasoning controls |
Deployment choice | Compare hosted service with feasible self-hosting |
Benchmarks and evidence
Moonshot's model card reports 65.7 on Terminal-Bench 4.0 and 92.7 on BenchCAD with tools. It also reports 61.2 on the internal Kimi Code Bench 3.0 Long-horizon evaluation. These results point to coding, tool-assisted spatial work and sustained execution as useful areas to test. Most results in the card use maximum reasoning effort, and some comparisons use different published harnesses or fallback settings. They are provider-reported evaluations, not an independently reproduced head-to-head test. Compare the same model configuration, tools and task budget before drawing conclusions about performance in your own workflow.
Evidence | Reported result | How to interpret it |
|---|---|---|
Terminal-Bench 4.0 | 65.7 | Moonshot model card; harness conditions apply |
BenchCAD with tools | 92.7 | Tool-assisted result, not text-only generation |
Kimi Code Bench 3.0 | 61.2 | Internal long-horizon benchmark |
Reasoning configuration | Usually max in model card | Do not compare with lower-effort settings silently |
Pricing and total cost
The Kimi API is billed by usage separately from Kimi memberships and Kimi Code plans. The rendered official pricing table lists standard K3.1 Pro at $2 input and $10 output per million tokens, with Highspeed at $4 and $20. Both list $0.20 cached input. Standard five-minute and one-hour cache writes are $2 and $4; Highspeed writes are $4 and $8. The default cache TTL is five minutes. K3.1 Pro batch inference is billed at 60% of standard price. At base standard rates, 1M uncached input plus 100K output costs $3, versus $6 for Highspeed, excluding other charges. Self-hosting costs are different again: weights do not make hardware, serving engineering or operations free.
Cost component | Published term or planning boundary |
|---|---|
Hosted API | Usage-based; separate from product subscriptions |
Standard input/output | $2 / $10 per million tokens |
Highspeed input/output | $4 / $20 per million tokens |
Cached input, both variants | $0.20 per million tokens |
Standard cache write 5m / 1h | $2 / $4 per million tokens |
Highspeed cache write 5m / 1h | $4 / $8 per million tokens |
Context caching | Repeated prefixes; default TTL documented as 5m |
Batch inference | 60% of standard rate for supported K3.1 Pro batch use |
Self-hosting | Hardware, serving and operations costs apply |
Taxes | Pricing documentation says applicable taxes are separate |
Limitations and deployment tradeoffs
Self-hosting a 2.8T-parameter model is a substantial infrastructure project, even with mixture-of-experts activation and quantization. The official license grants broad rights but requires a separate agreement for Model-as-a-Service businesses whose aggregate affiliate revenue exceeds $20 million over a consecutive 12-month period. It requires prominent Kimi K3.1-Pro attribution for commercial products above 100 million monthly active users or $20 million in monthly revenue. It also states exceptions for internal use and official or certified inference channels. Read the license itself rather than relying on a summary for eligibility. In evaluation, separate provider demonstrations from repeatable workload performance and keep the standard and Highspeed variants distinct in cost and latency reports.
Consideration | What to check |
|---|---|
Custom license | Review Model-as-a-Service and large-product conditions |
Infrastructure scale | Do not infer laptop deployment from open weights |
Hosted versus local | Account for different tools, latency and support |
Benchmark configurations | Preserve effort, harness and fallback details |
Variant pricing | Highspeed costs twice standard at base input/output rates |
How to evaluate Kimi K3.1 Pro for your workload
Start with a small, representative set of open-weight multimodal engineering tasks, including ordinary cases, difficult edge cases and tasks with an objectively checkable answer. For this model, a useful pilot includes a code-and-image task, a parametric modeling task and a sourced financial analysis. Keep the prompts, available tools, source material and acceptance criteria consistent across candidates. Record the exact model ID, reasoning configuration, date and any provider-specific access program so the comparison can be repeated later.
Measure completed-task quality rather than the fluency of the response. Ask a reviewer to identify unsupported claims, incorrect changes, missing evidence and avoidable manual repair. An agent that stops honestly when evidence is insufficient can be more useful than one that produces an impressive-looking but invalid artifact. Where tools are involved, track invalid arguments, failed calls and recovery behavior separately from the final answer. A single successful demonstration is a reason to continue testing, not proof of a dependable service level.
Pilot measure | What to record |
|---|---|
Task completion | Was the requested artifact actually completed and checked? |
Evidence quality | Can a reviewer trace decisive claims to sources or tests? |
Correction burden | How much expert time was needed after the model finished? |
Latency | Measure end-to-end time, not just first-token speed |
Total cost | Include tokens, tools, retries and human review |
Failure handling | Did it report uncertainty and recover without inventing success? |
Use a staged rollout only after the pilot meets the target you set in advance. Preserve a known baseline and a way to compare new requests with the previous configuration. If the model, tool interface, access policy or price changes, rerun the affected cases instead of assuming an old result still applies. The purpose of evaluation is a defensible choice for your actual work, not a claim that one model is universally best.
Who should use it, and what should you compare?
K3.1 Pro fits teams evaluating a capable open-weight option for advanced engineering and multimodal work, particularly when deployment control matters. Hosted API access is the simpler starting point for many teams; self-hosting only makes sense after infrastructure and license requirements are understood. A financial-research workflow also needs reliable retrieval, calculation checks and traceable sources rather than trusting a fluent narrative. If latency is the deciding factor, compare Highspeed on the same tasks instead of assuming the stated speed multiplier translates directly into user-visible savings. Keep a strong hosted frontier baseline in the pilot.
Alternative | Why compare it | Decision boundary |
|---|---|---|
Kimi K3 | Previous-generation comparison | Keep version-specific evidence separate |
Claude Opus 5.5 | Hosted coding and knowledge work | Compare completed artifacts and total cost |
GPT-6 Astra | Frontier agentic baseline | Use the same tasks and tool permissions |
DeepSeek V4.1 Flash | Alternative open-weight shortlist | Check its own license and deployment requirements |
Frequently asked questions
Are the K3.1 Pro weights actually available?
Yes. The official Moonshot repository provides the model card and weights. This is stronger evidence than the launch's earlier promise to release them, but it does not eliminate infrastructure or licensing requirements.
Is K3.1 Pro unrestricted open source?
No. Open-weight is the safer description. Its custom license includes conditions for certain Model-as-a-Service businesses and large commercial products, with exceptions defined in the license. Read the exact terms for your deployment.
Is Highspeed the same API model ID?
No. The model list distinguishes kimi-k3.1-pro from kimi-k3.1-pro-highspeed. Moonshot describes the latter as the same intelligence at roughly two to three times the speed; verify its pricing and access separately.
Does Kimi Code membership cover API usage?
The pricing documentation distinguishes API pay-as-you-go billing from Kimi membership and Kimi Code products. Check the plan and endpoint you actually use rather than assuming one product's subscription includes another service.
Related model guides
Compare the full AI model directory
Sources and editorial method
Official model card and weights
Specifications, access conditions and prices were checked against the linked provider sources on November 10, 2026. Provider-reported benchmarks are attributed and are not presented as independently reproduced tests. Missing specifications remain unverified rather than being inferred from a related model. Practical workflow advice is an evaluation recommendation, not a claim that we ran new hands-on trials. Recheck official terms before purchase or deployment.