Kimi K3.1 Pro
Moonshot AI
- Released
- October 23, 2026
- Data date
- October 24, 2026
- Predecessor
- Kimi K3
- Benchmark-Meta-Score
- UnavailableNo independent results
- Price
- Input · Output / 1M tokens
- Input
- Text, images, and video. The API guide shows video input through an uploaded file.
- Output
- Text
Kimi K3.1 Pro is Moonshot's reasoning model. Its context window reaches 1,048,576 tokens on the Pro or Allegretto Kimi Code tier; lower tiers have a smaller limit. Kimi Code requires Thinking and offers Low, High, and Max effort. Moonshot says the Highspeed variant has the same model quality, so it is not a separate model profile.
The API guide documents text, image, and video input. max_completion_tokens defaults to 131,072 and can be set up to 1,048,576. This API value shares the 1,048,576-token context window with the input.
Moonshot announced a release of the full weights by November 4, 2026. As of October 24, neither the weights nor their license had been published.
Specifications and access
| Specification | Value and source |
|---|---|
| Provider model IDs | API: kimi-k3.1-pro. Kimi Code: k3.1-pro and k3.1-pro-highspeed.Source |
| Model variant | Kimi K3.1 Pro is the first model in the K3.1 series. Highspeed is a speed variant with the same model quality according to Moonshot, not a separate model profile.Source |
| Highspeed variant | Kimi Code lists roughly two to three times the speed and about twice the quota usage. The API platform says approximately three times the speed.Source |
| Kimi Code subscription context | Kimi Code offers K3.1 Pro from Plus or Moderato. The 1,048,576-token context requires Pro or Allegretto.Source |
| Announced weight release | Moonshot says the full weights will be released by November 4, 2026. The license and actual release remained unknown on the verification date.Source |
| Parameter figures | Moonshot says 3T-class. Artificial Analysis lists 2,800 billion total and 104 billion active parameters; the more precise figure is not from a published Moonshot model card.Source |
| Input modalities | Text, images, and video. The API guide shows video input through an uploaded file.Source |
| Output modalities | TextSource |
| API default output limit | The model-specific Kimi K3.1 Pro quickstart gives a default max_completion_tokens value of 131,072 tokens.Source |
| API maximum output limit | The model-specific Kimi K3.1 Pro quickstart allows max_completion_tokens up to 1,048,576. Actual output shares the 1,048,576-token context window with the input.Source |
| Reasoning | Thinking is mandatory for Kimi Code models. Low, High, and Max effort are documented; Max is the default.Source |
| Release | 2026-10-23Source |
| Provider model ID | kimi-k3.1-pro, k3.1-pro, k3.1-pro-highspeedSource |
| Context window | 1,048,576 tokensSource |
| Total parameter class | 3T-classSource |
| Input modalities | text, image (The launch article names text and images. This is a partial source listing; the model-specific API quickstart additionally documents video_url input.)Source |
| Output modalities | textSource |
| Reasoning modes | Low, High, MaxSource |
| Highspeed variant | same model quality; approximately 2–3x speed and about 2x quota usageSource |
| Kimi Code access | Plus/Moderato or aboveSource |
| Model weights status | announced, not yet releasedSource |
| CNY input price per million tokens | 13Source |
| CNY output price per million tokens | 68Source |
| CNY cache write price per million tokens | 13Source |
| CNY cache-hit price per million tokens | 1.3Source |
| Highspeed prices in CNY per million tokens | CNY 26 input; CNY 136 output; CNY 26 cache write; CNY 1.30 cache hitSource |
| API default output limit | Model-specific max_completion_tokens default: 131,072 tokensSource |
| Maximum configurable API output limit | Model-specific API request ceiling for max_completion_tokens: 1,048,576 tokens. Output shares the context window with input.Source |
| Input | Text, image, and videoSource |
| Provider model ID | OpenRouter model ID: moonshotai/kimi-k3.1-pro
Hugging Face bridge ID listed by OpenRouter catalog: not provided (null)Source |
| Deployment options | Moonshot AI [moonshotai/mxfp4]; endpoint name=Moonshot AI | moonshotai/kimi-k3.1-pro-20261023; quantization=mxfp4; raw status code=0
Moonshot AI [moonshotai/highspeed; service_tier=priority]; endpoint name=Moonshot AI | moonshotai/kimi-k3.1-pro-20261023; quantization=mxfp4; raw status code=0 (This lists every provider row for the exact OpenRouter route(s) queried. Status codes are preserved as returned, not interpreted or used to filter offers.)Source |
| Context values by source | Moonshot AI [moonshotai/mxfp4]; context_length=1048576 tokens
Moonshot AI [moonshotai/highspeed; service_tier=priority]; context_length=1048576 tokens (These are limits per OpenRouter provider route. They do not replace separately documented direct-API or model-configuration values.)Source |
| Source output limit | Moonshot AI [moonshotai/mxfp4]; max_completion_tokens=943718 tokens
Moonshot AI [moonshotai/highspeed; service_tier=priority]; max_completion_tokens=943718 tokens (These are output limits for the corresponding OpenRouter provider route, not direct-API limits. Provider differences remain separate.)Source |
| API pricing in USD per million tokens | Moonshot AI [moonshotai/mxfp4]; prompt USD/1M=2; completion USD/1M=10; cache-read USD/1M=0.2; raw discount field=0
Moonshot AI [moonshotai/highspeed; service_tier=priority]; prompt USD/1M=4; completion USD/1M=20; cache-read USD/1M=0.2; raw discount field=0 (Per-token API prices were multiplied exactly by 1,000,000. Raw discount fields and returned pricing schedules are preserved separately. No endpoint pricing object in these returned rows contained an input_cache_write/cache-write price; it is not reported as zero.)Source |
| API capabilities | Moonshot AI [moonshotai/mxfp4]; supported_parameters=["reasoning","include_reasoning","max_tokens","stop","frequency_penalty","presence_penalty","structured_outputs","response_format","tool_choice","tools","reasoning_effort"]; supports_tool_choice={"auto":true,"function":false,"none":true,"required":true}; supports_implicit_caching=true; native_tools={}; supports_voice_cloning=false; supports_multiple_audio_references=false; supports_image_reference=false; supports_diarization=false
Moonshot AI [moonshotai/highspeed; service_tier=priority]; supported_parameters=["reasoning","include_reasoning","max_tokens","stop","frequency_penalty","presence_penalty","structured_outputs","response_format","tool_choice","tools","reasoning_effort"]; supports_tool_choice={"auto":true,"function":false,"none":true,"required":true}; supports_implicit_caching=true; native_tools={}; supports_voice_cloning=false; supports_multiple_audio_references=false; supports_image_reference=false; supports_diarization=false (Supported parameters and returned API capability flags are listed per provider route. The route-level architecture returned in the same response is: {"input_modalities":["text","image","video"],"instruct_type":null,"modality":"text+image+video->text","output_modalities":["text"],"tokenizer":"Other"}.)Source |
| Quantized variant | Moonshot AI [moonshotai/mxfp4]; tag=moonshotai/mxfp4; quantization=mxfp4
Moonshot AI [moonshotai/highspeed; service_tier=priority]; tag=moonshotai/highspeed; quantization=mxfp4 (Tags and quantization values are taken literally from endpoint responses. Missing values are not inferred from names.)Source |
Benchmark meta-analysis
Overall and area scores
These Gradually indexes place selected published measurements on a 0 to 100 scale. They are not accuracy percentages and do not guarantee results in your application.
Overall index
Not enough independent results for a score yet
provisional
- Test series
- 0
- Test families
- 0 / 65
- Sources
- 0
- Share of the fixed test panel (informative)
- 0 %
Documented method · Included test families and measurements
Method version: benchmark-meta-score-v3
The index includes active measurements from reviewed primary sources when at least five models are compared in the same documented comparison slice across dataset, harness, agent, tools, and trial count. Each benchmark series and evaluator contributes one shared setup slice from the newest capture. When one source and series include several explicitly documented reasoning tiers, the highest documented tier (none through max) is selected, never the best result value. The explicit source label “Non-reasoning” is treated as none. Unknown or ambiguous tiers exclude the model from that series. Within each series, percentages and rates are logit-transformed, other units stay unchanged, and all values are then standardized. A one-factor model then estimates one shared strength per model and, per series, how sharply it separates the models. Area scores move away from that strength only as far as the model's own results in the area support. The overall score weights the core areas at 75% and finance, law and health at 25%. It requires at least eight test series from two sources and three areas. The 0 to 100 scale refers to a fixed reference group of 29 language models whose average sits at 50.
Published model-specific parameters can differ. This is not a controlled experiment; comparisons with such differences are marked provisional.
Unmeasured series are not entered as zero. Where a model has no own results in a core area, its overall strength stands in, and the note “estimated without own results in …” names the area. Fewer test series lead to a larger standard error.
The 95% interval follows from the standard error of 200 recalculations with randomly redrawn test families. It describes how precisely the recorded test series determine the value. It is no prediction for future tasks or for later recalculations with other participants.
The source count is not a count of independent experiments.
OECD handbook on composite indicatorsAll 0 test families and 0 of 0 raw measurements are listed here. The complete calculation for every model is available in the JSON export.
Benchmark sources at a glance
Each evaluator measures different capabilities under its own conditions. Test series, composite indices and their components stay separate. The detailed results below retain their scales, configurations and source dates.
| Evaluator | Test series | Composite indices | Recorded active measurements |
|---|---|---|---|
| Artificial Analysis | 14 | 6 | 147 |
| LMArena | 0 | 0 | 4 |
Coverage is not a quality ranking. Overlapping tests and multiple settings do not count as independent experiments. Vendor claims, historical and disputed results are excluded from this overview and remain labelled in the detailed evidence.
126 measurements from these sources still lack a verified test-family identity and do not increase the test-series count.
Page 1 of 22
Measurements without matching peer values
These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.
Terminal-Bench 4.0 (Show measurement, test conditions, and source)
- Source value
- 65.7
- Score
- 65.7
- Metric
- percent
- Unit
- %
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Moonshot, Kimi K3.1 Pro benchmark table
- Status
- active
- Retrieved at
- 2026-10-24
- Methodology
- Vendor-reported result. Claude Code harness; maximum reasoning effort.
- Context
- The full source table lists this score, but no exact sample size is disclosed.
Terminal-Bench Science 0.1 (Show measurement, test conditions, and source)
- Source value
- 38.1
- Score
- 38.1
- Metric
- percent
- Unit
- %
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Moonshot, Kimi K3.1 Pro benchmark table
- Status
- active
- Retrieved at
- 2026-10-24
- Methodology
- Vendor-reported result. Claude Code harness; maximum reasoning effort. Moonshot says its internal run gave a different result for the named DeepSeek comparison, but the Kimi score here is the Kimi run.
- Context
- The full source table lists this score, but no exact sample size is disclosed.
SWE Marathon v1.1 (Show measurement, test conditions, and source)
- Source value
- 51.7
- Score
- 51.7
- Metric
- percent
- Unit
- %
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Moonshot, Kimi K3.1 Pro benchmark table
- Status
- active
- Retrieved at
- 2026-10-24
- Methodology
- Vendor-reported result. Claude Code harness; maximum reasoning effort.
- Context
- The full source table lists this score, but no exact sample size is disclosed.
BenchCAD w/o tools (Show measurement, test conditions, and source)
- Source value
- 72.8
- Score
- 72.8
- Metric
- Voxel IoU
- Unit
- %
- Direction
- Higher is better
- Sample
- 1000
- Source type
- vendor-reported
- Evaluator
- Moonshot, Kimi K3.1 Pro benchmark table
- Status
- active
- Retrieved at
- 2026-10-24
- Methodology
- Vendor-reported result. Random 1,000-case sample from 17,900 Vision2Code examples; maximum reasoning; single-shot generation with official prompts and scoring pipeline.
BenchCAD w/ tools (Show measurement, test conditions, and source)
- Source value
- 92.7
- Score
- 92.7
- Metric
- Voxel IoU
- Unit
- %
- Direction
- Higher is better
- Sample
- 1000
- Source type
- vendor-reported
- Evaluator
- Moonshot, Kimi K3.1 Pro benchmark table
- Status
- active
- Retrieved at
- 2026-10-24
- Methodology
- Vendor-reported result. Random 1,000-case sample from 17,900 Vision2Code examples; maximum reasoning; Claude Code harness.
Program Bench (Show measurement, test conditions, and source)
- Source value
- 31.5
- Score
- 31.5
- Metric
- percent
- Unit
- %
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Moonshot, Kimi K3.1 Pro benchmark table
- Status
- active
- Retrieved at
- 2026-10-24
- Methodology
- Vendor-reported result. Claude Code harness; maximum reasoning; binary task score requires the result to meet at least 95% of task requirements.
- Context
- The full source table lists this score, but no exact sample size is disclosed.
DeepSWE v1.1 (Show measurement, test conditions, and source)
- Source value
- 75.5
- Score
- 75.5
- Metric
- percent
- Unit
- %
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Moonshot, Kimi K3.1 Pro benchmark table
- Status
- active
- Retrieved at
- 2026-10-24
- Methodology
- Vendor-reported result. Claude Code harness; maximum reasoning.
- Context
- The full source table lists this score, but no exact sample size is disclosed.
PostTrain Bench v1.2 (Show measurement, test conditions, and source)
- Source value
- 36.5
- Score
- 36.5
- Metric
- percent
- Unit
- %
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Moonshot, Kimi K3.1 Pro benchmark table
- Status
- active
- Retrieved at
- 2026-10-24
- Methodology
- Vendor-reported result. Internal run using Claude Code and H20 GPUs; judge was DeepSeek V4.1 Flash 1017 at maximum effort.
- Context
- The full source table lists this score, but no exact sample size is disclosed.
Kimi Code Bench 2.0 (Internal) (Show measurement, test conditions, and source)
- Source value
- 77.1
- Score
- 77.1
- Metric
- percent
- Unit
- %
- Category
- Coding; internal benchmark, not comparison eligible
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Moonshot, Kimi K3.1 Pro benchmark table
- Status
- active
- Retrieved at
- 2026-10-24
- Methodology
- Vendor-reported result. Claude Code harness; maximum reasoning.
Kimi Code Bench 3.0 Long-horizon (Internal) (Show measurement, test conditions, and source)
- Source value
- 61.2
- Score
- 61.2
- Metric
- percent
- Unit
- %
- Category
- Coding; internal benchmark, not comparison eligible
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Moonshot, Kimi K3.1 Pro benchmark table
- Status
- active
- Retrieved at
- 2026-10-24
- Methodology
- Vendor-reported result. Claude Code harness; maximum reasoning.
HLE-Full w/ tools (Show measurement, test conditions, and source)
- Source value
- 70
- Score
- 70
- Metric
- percent
- Unit
- %
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Moonshot, Kimi K3.1 Pro benchmark table
- Status
- active
- Retrieved at
- 2026-10-24
- Methodology
- Vendor-reported result. Tool-assisted Kimi Code evaluation; maximum reasoning.
- Context
- The full source table lists this score, but no exact sample size is disclosed.
BrowseComp (Show measurement, test conditions, and source)
- Source value
- 92.1
- Score
- 92.1
- Metric
- percent
- Unit
- %
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Moonshot, Kimi K3.1 Pro benchmark table
- Status
- active
- Retrieved at
- 2026-10-24
- Methodology
- Vendor-reported result. Context compaction at 200,000 tokens; up to 500 steps; temperature 1.0 and top-p 1.0.
- Context
- The full source table lists this score, but no exact sample size is disclosed.
Page 1 of 22
Published prices
- API price listed by Artificial Analysis
- $2 input, $10 output, and $0.20 cached input per 1 million tokens. Kimi's official API publishes CNY rates separately.Source
- Official API pricing
- K3.1 Pro: CNY 13 input, CNY 68 output, CNY 13 cache write, and CNY 1.30 cache hit per million tokens.Source
- Highspeed API pricing
- CNY 26 input, CNY 136 output, CNY 26 cache write, and CNY 1.30 cache hit per million tokens.Source
Predecessor and successor
Sources and data date
| Type | Evidence and data date |
|---|---|
| Price | Base API prices without caching or batch discounts. Higher context tiers and other rates are listed under Costs. |
| Input | Documented modalities. Retrieved October 24, 2026. |
| Output | Documented modalities. Retrieved October 24, 2026. |
| Research status | Research date October 24, 2026. 162 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Not published | Model weights license Note Moonshot announced the full weights for release by November 4, 2026, but the weights and license were not yet published on October 24. |
| Unresolved | Independently confirmed parameter count Note Artificial Analysis lists 2,800B total and 104B active, while Moonshot describes the model only as 3T-class; Moonshot has not published an exact model card parameter count. |
| Not published | Architecture and MoE Note The checked Moonshot release, model-specific API guide, and Kimi Code documents do not identify the model architecture or say whether it is MoE. |
| Not published | Knowledge cutoff Note No training-data or knowledge cutoff is stated in the checked exact-model release and API sources. |
| Unresolved | Current OpenCompass results Note The current public OpenCompass results could not be fully retrieved during the source check. Possible results for this exact model therefore remain unresolved. |
| Not found in the inspected sources | Exact-model ARC Prize results Note The full current public ARC Prize evaluations and model registry were read; no exact identity for this target appears across ARC-AGI-1/2/3. Similar names without an exact model identity bridge were not mapped. |
| Not found in the inspected sources | Exact-model Scale Labs results Note The Scale Labs index and current public result boards in the bounded capture set were read; no exact row for this target was found. Kimi-K3 and undated DeepSeek V4.1 Flash rows are not bridged to newer targets. |
| Additional source | Moonshot Kimi K3.1 Pro announcement (retrieved October 24, 2026) · Moonshot Kimi K3.1 Pro announcement · Editorial description reviewed October 24, 2026 |
| Additional source | Kimi Code model documentation (retrieved October 24, 2026) · Kimi Code model documentation · Editorial description reviewed October 24, 2026 |
| Additional source | Kimi Code release notes (retrieved October 24, 2026) · Kimi Code release notes · Editorial description reviewed October 24, 2026 |
| Additional source | Kimi API platform (retrieved October 24, 2026) |
| Additional source | LiveBench (retrieved October 24, 2026) |
| Additional source | LiveBench (retrieved October 24, 2026) |
| Additional source | LiveBench (retrieved October 24, 2026) |
| Additional source | LMArena (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |