Kimi K2 Instruct 0905
Kimi K2 0905 is the September update of Kimi K2 0711. It is a large-scale Mixture-of-Experts (MoE) language model developed by Moonshot AI, featuring 1 trillion total parameters with 32 billion active per forward pass. It supports long-context inference up to 256k tokens, extended from the previous 128k.
This update improves agentic coding with higher accuracy and better generalization across scaffolds, and enhances frontend coding with more aesthetic and functional outputs for web, 3D, and related tasks. Kimi K2 is optimized for agentic capabilities, including advanced tool use, reasoning, and code synthesis. It excels across coding (LiveCodeBench, SWE-bench), reasoning (ZebraLogic, GPQA), and tool-use (Tau2, AceBench) benchmarks. The model is trained with a novel stack incorporating the MuonClip optimizer for stable large-scale MoE training.
- Organization
- Moonshot AI
- Family
- kimi-k2
- Providers
- 5
- Context
- 262,144
- Output limit
- 98,304
- Knowledge
- —
- Release
- 2025-09-05
- Updated
- 2025-09-05
- Weights
- Open
- Input
- text
- Output
- text
- Capabilities
- Tools, Temperature
Providers
Compare every cataloged offer for this model, including its provider, route, token limits, price, and capabilities.
| Provider | Model ID | Context | Output | Input / output · 1M | Reasoning | Tools | Structured | Details |
|---|---|---|---|---|---|---|---|---|
| AI/ML API | moonshotai/kimi-k2-0905 | 262,144 | 98,304 | $1.3754 / $4.1262 | No | Yes | — | View |
| Kilo Gateway | moonshotai/kimi-k2-0905 | 262,144 | 98,304 | $0.60 / $2.50 | No | Yes | Yes | View |
| Novita AI | moonshotai/kimi-k2-0905 | 262,144 | 98,304 | $0.60 / $2.50 | No | Yes | Yes | View |
| Novitavia OpenRouter | moonshotai/kimi-k2-0905 fp8 | 262,144 | 98,304 | $0.60 / $2.50 | No | Yes | Yes | View |
| Prime Inference | moonshotai/kimi-k2-0905 | 262,144 | 98,304 | $0.60 / $2.50 | No | Yes | Yes | View |
Prices and limits apply to each offer. A dash means the value is not provided. OpenRouter offers show the hosting provider and routing channel separately.
Pricing
Compare token rates and cache charges across available routes and hosting configurations.
Lowest listed input and output prices across the offers above; they may belong to different providers. These are base rates, before conditional pricing, discounts or additional fees.
moonshotai/kimi-k2-0905Observed - Input
- $0.60
- Output
- $2.50
Hosting prices through OpenRouter All endpoints and cache rates
| Endpoint | Input / 1M | Output / 1M | Cache read / 1M | Cache write / 1M |
|---|---|---|---|---|
| Novitanovita/fp8 | $0.60 | $2.50 | — | — |
Each row is a hosting configuration. Open an offer in Providers for its full pricing conditions and additional fees. Endpoint observations may differ from the routing catalog's base price.
Performance
Review reported latency and output-throughput percentiles from the latest observed rolling window.
Time to first token · milliseconds
All latency percentiles 1 endpoints
| Endpoint | P50 | P75 | P90 | P99 |
|---|---|---|---|---|
| Novitanovita/fp8 | 667 ms | 932 ms | 1,710.5 ms | 9,811.3 ms |
Output throughput · tokens per second
All throughput percentiles 1 endpoints
| Endpoint | P50 | P75 | P90 | P99 |
|---|---|---|---|---|
| Novitanovita/fp8 | 40 t/s | 44 t/s | 49 t/s | 125.43 t/s |
Percentiles describe the observed request distribution. These are snapshots of a rolling window, not a historical time series.
Uptime
Compare successful request rates across the latest 5-minute, 30-minute, and 24-hour observations.
Reported by OpenRouter at the time of collection; rate-limited requests are excluded. A dash means no measurement was reported. This is not a live availability check.
Benchmarks
Review attributed evaluation results on each source's original scale.
moonshotai/kimi-k2-0905Observed | Design Arena | Category | Elo | Win rate | Rank |
|---|---|---|---|---|
| models | codecategories | 1,102 | 48.5% | 96 |
| models | website | 1,112 | 48.3% | 99 |
Sources: Artificial Analysis and Design Arena, via OpenRouter. Scores retain their original scales and are not OpenAI Suite ratings.
Detailed evaluations
| Evaluation | Configuration | Source | Results | Details |
|---|---|---|---|---|
| models · website | Kimi K2 Instruct 0905moonshotai/kimi-k2-0905 | Design Arena | 1,090 Elo | Details |
| models · codecategories | Kimi K2 0905 Previewmoonshotai/kimi-k2-0905 | Design Arena | 1,100 Elo | Details |
| gpqa diamond | MoonshotAI: Kimi K2 0905moonshotai/kimi-k2-0905 | OpenRouter | 74.663% | Details |
| tau bench verified airline | MoonshotAI: Kimi K2 0905moonshotai/kimi-k2-0905 | OpenRouter | 57.207% | Details |
| models · website | Kimi K2 0905 Previewmoonshotai/kimi-k2-0905 | Design Arena | 1,109 Elo | Details |
| Composite indices | Kimi K2 0905moonshotai/kimi-k2-0905 | Artificial Analysis | Intelligence 15.3 · Coding — · Agentic — | Details |
Sources: Artificial Analysis, Design Arena and OpenRouter. Evaluation configurations and scales differ; results are not interchangeable. Updated Oct 15, 2026.
Technical details & API
Explore specifications, supported parameters, reasoning controls, and API identifiers for each OpenRouter route.
moonshotai/kimi-k2-0905Observed - Context
- 262,144
- Maximum output
- 98,304
- Tokenizer
- Other
- Input
- text
- Output
- text
- Moderation
- No
- Knowledge cutoff
- 2024-12-31
- Added to OpenRouter
- 2025-09-04
https://openrouter.ai/api/v1moonshotai/kimi-k2-0905moonshotai/kimi-k2-0905moonshotai/Kimi-K2-Instruct-0905Supported parameters
frequency_penaltymax_tokenspresence_penaltyrepetition_penaltyresponse_formatseedstopstructured_outputstemperaturetool_choicetoolstop_ktop_p
Frequently asked questions
Answers based on the model metadata and serving offers currently in the catalog.
How much does Kimi K2 Instruct 0905 cost?
Listed rates start at $0.60 per million input tokens and $2.50 per million output tokens. Prices vary by provider and configuration; see Pricing for details.
Which providers offer Kimi K2 Instruct 0905?
There are 5 cataloged offers across 5 providers. Compare model IDs, prices and limits in Providers.
What is the context window?
The cataloged model context is 262,144 tokens. Each serving endpoint may apply a different limit.
Does it support tools and structured output?
Tool calling: Yes. Structured output: Not reported. Support can vary by endpoint.
Resources
Modified MITCatalog history · 1 recorded revision
Metadata observations since this model was first cataloged. These are not model release versions.
- ·
174f485f1ff6