概要
Kimi K2.5 Instant is the fast-response variant of Moonshot AI's flagship open-weight model, a 1-trillion-parameter Mixture-of-Experts architecture that activates only 32 billion parameters per token. Released January 27, 2026, it represents the first Moonshot model to offer native multimodal (image and video) inputs alongside text, making it the leading open-weight model with integrated vision capabilities. The Instant variant specifically disables chain-of-thought reasoning to prioritize speed and cost efficiency, targeting interactive applications like chatbots, quick code generation, and real-time assistants.
The model's most distinctive innovation is Agent Swarm — a capability unique among frontier models — which orchestrates up to 100 parallel sub-agents using Parallel-Agent Reinforcement Learning (PARL). This enables ~4.5× speedups on parallelizable research and analysis tasks. Combined with aggressive API pricing ($0.60 per million input tokens, roughly 25× cheaper than Claude Opus 4.5), Kimi K2.5 Instant occupies a compelling niche for teams needing frontier-class reasoning at open-weight economics.
However, the model carries real trade-offs. Output generation is extremely verbose (~89 million tokens across Artificial Analysis's evaluation suite, roughly 6× the median model), which partially erodes the cost advantage. Self-hosting requires substantial GPU infrastructure (~630GB for INT4 weights, 8× high-end GPUs), agent workflows suffer ~12% tool-call failure rates, and English writing quality lags behind Claude and GPT. The 256K context window, while large, trails Gemini's 1M window by 4×. For teams willing to manage these constraints — particularly those building agentic or multimodal pipelines — K2.5 Instant offers capabilities unavailable from any competing open-weight model.
ベンチマーク＆性能
Kimi K2.5's benchmark performance spans reasoning, vision, coding, and agentic domains. The following table summarizes key scores from the official technical report. Note: most published benchmarks use the Thinking variant; the Instant variant trades reasoning depth for speed.
| Benchmark | Kimi K2.5 (Thinking) | GPT-5.2 (xhigh) | Claude Opus 4.5 | Gemini 3 Pro | DeepSeek V3.2 |
|---|---|---|---|---|---|
| **Reasoning & Knowledge** |||||
| HLE-Full (w/ tools) | 50.2 | 45.5 | 43.2 | 45.8 | 40.8† |
| AIME 2025 (avg@32) | 96.1 | 100 | 92.8 | 95.0 | 93.1 |
| HMMT 2025 Feb (avg@32) | 95.4 | 99.4 | 92.9* | 97.3* | 92.5 |
| GPQA-Diamond (avg@8) | 87.6 | 92.4 | 87.0 | 91.9 | 82.4 |
| MMLU-Pro | 87.1 | 86.7* | 89.3* | 90.1 | 85.0 |
| **Vision & Multimodal** |||||
| MMMU-Pro | 78.5 | 79.5* | 74.0 | 81.0 | — |
| MathVision | 84.2 | 83.0 | 77.1* | 86.1* | — |
| OCRBench | 92.3 | 80.7* | 86.5* | 90.3* | — |
| VideoMMMU | 86.6 | 85.9 | 84.4* | 87.6 | — |
| LongVideoBench | 79.8 | 76.5* | 67.2* | 77.7* | — |
| **Coding** |||||
| SWE-Bench Verified | 76.8 | 80.0 | 80.9 | 76.2 | 73.1 |
| SWE-Bench Multilingual | 73.0 | 72.0 | 77.5 | 65.0 | 70.2 |
| LiveCodeBench v6 | 85.0 | — | 82.2* | 87.4* | 83.3 |
| Terminal Bench 2.0 | 50.8 | 54.0 | 59.3 | 54.2 | 46.4 |
| **Agentic Search** |||||
| BrowseComp (Agent Swarm) | 78.4 | — | — | — | — |
| BrowseComp (w/ ctx mgmt) | 74.9 | 57.8 | 59.2 | 67.6 | — |
| DeepSearchQA | 77.1 | 71.3* | 76.1* | 63.2* | 60.9* |
| WideSearch (Agent Swarm) | 79.0 | — | — | — | — |
| Seal-0 | 57.4 | 45.0 | 47.7* | 45.5* | 49.5* |
| **Long Context** |||||
| Longbench v2 | 61.0 | 54.5* | 64.4* | 68.2* | 59.8* |
*Scores marked with * were re-evaluated by Moonshot under identical conditions. †DeepSeek V3.2 HLE score is text-only subset.*
**Key observations:**
- **Agentic search dominance**: K2.5 leads all competitors on BrowseComp (74.9% single-agent, 78.4% swarm) and DeepSearchQA (77.1%). The Agent Swarm mode has no equivalent in competing models.
- **Vision leadership among open-weights**: At 92.3% on OCRBench and 92.6% on InfoVQA, K2.5 surpasses GPT-5.2 and Claude on document/vision benchmarks.
- **Mathematical reasoning**: 96.1% on AIME 2025 places it within 4 points of GPT-5.2's perfect score, and ahead of Claude (92.8%) and Gemini (95.0%).
- **Coding gap**: SWE-Bench Verified (76.8%) trails Claude Opus 4.5 (80.9%) and GPT-5.2 (80.0%), though it leads DeepSeek V3.2 (73.1%) and matches Gemini (76.2%).
- **LM Arena Code Elo**: The Instant variant scores 1526 (#10), while the Thinking variant scores 1550 (#7), as of February 2026 snapshots from Propel Code.
詳細比較
### Kimi K2.5 Instant vs GPT-5.2
GPT-5.2 leads on pure reasoning benchmarks (GPQA: 92.4% vs 87.6%, AIME: 100% vs 96.1%) and has a more mature consumer ecosystem (plugins, voice, image generation). However, K2.5 Instant is ~4× cheaper on input tokens ($0.60 vs $2.50/1M), excels on agentic search tasks (BrowseComp: 74.9% vs 57.8% without swarm), and offers open-weight deployment. GPT-5.2's context window is 128K vs K2.5's 256K. For teams building autonomous research or document-heavy pipelines, K2.5 offers dramatically better economics; for polished consumer products, GPT-5.2 remains stronger.
### Kimi K2.5 Instant vs Claude Opus 4.5
Claude Opus 4.5 wins on code quality (SWE-Bench: 80.9% vs 76.8%) and English writing naturalness, with a more mature agentic tooling ecosystem (Claude Code, Agent Teams). However, Claude costs 25× more on input tokens ($15.00 vs $0.60/1M). K2.5's vision capabilities are native and broader (video support), while Claude's 200K context is slightly smaller. Claude has no equivalent to Agent Swarm. For coding-heavy teams on a budget, K2.5 is compelling; for polished English output and critical code review, Claude remains superior.
### Kimi K2.5 Instant vs Gemini 3 Pro
Gemini 3 Pro dominates on context length (1M vs 256K — a 4× advantage) and leads on several vision benchmarks (MMMU-Pro: 81.0% vs 78.5%, MathVision: 86.1% vs 84.2%). Gemini is also ~2× cheaper on input ($1.25 vs $0.60, but significantly cheaper on output). K2.5 counters with open-weight availability, Agent Swarm (no Gemini equivalent), stronger agentic search (BrowseComp swarm: 78.4%), and better OCR (92.3% vs 90.3%). For massive document ingestion, Gemini's context window wins; for autonomous multi-agent workflows, K2.5 is unmatched.
### Kimi K2.5 Instant vs DeepSeek V3.2
Both are open-weight MoE models from Chinese labs. DeepSeek V3.2 is significantly cheaper ($0.27/1M input vs $0.60) and uses fewer reasoning tokens. K2.5 leads on nearly every benchmark: AIME (96.1% vs 93.1%), SWE-Bench (76.8% vs 73.1%), GPQA (87.6% vs 82.4%), and especially agentic search (BrowseComp: 74.9% vs 51.4%). K2.5 also adds native multimodal support that DeepSeek V3.2 lacks. DeepSeek wins purely on cost efficiency; K2.5 wins on capability breadth.
コミュニティ評価
Developer and researcher sentiment toward Kimi K2.5 has been notably positive but calibrated. Several key themes emerge from the provided sources:
**Enthusiasm for Agent Swarm**: The Agent Swarm capability has generated the most excitement. Multiple reviewers describe it as a genuinely novel capability with no equivalent in competing models. BinaryVerse AI calls it 'the headline feature that differentiates this large language model from competitors,' while OpenAIToolsHub rates the model 8.0/10, noting 'K2.5 introduces a genuinely novel capability in Agent Swarm that no competitor matches.' Real-world testing by OpenAIToolsHub confirmed 4× speedups on parallelizable tasks like competitive analysis and codebase review.
**Coding community adoption**: Propel Code reports that K2.5 was added to LM Arena's Code Arena within days of release (February 2, 2026) and rapidly entered the top 10. The model is described as 'a serious coding model candidate, not just an experimental option.' Multiple sources highlight its particular strength in front-end/UI code generation from visual inputs, with one reviewer noting it produces UI code with 'sensible spacing, readable typography, and components that behave like real products.'
**Cost-benefit pragmatism**: Artificial Analysis notes K2.5's GDPval-AA Elo of 1309 places it 'behind only OpenAI and Anthropic models,' and its cost to run the Intelligence Index ($371) is 'more than 4x cheaper than Claude Opus 4.5 and GPT-5.2.' This price-performance ratio has attracted attention from startups and high-volume application builders. However, the verbosity issue is widely noted — Artificial Analysis flagged ~89M tokens generated during their evaluation, roughly 6× the median.
**Cautious deployment advice**: OraCore.dev describes agent runs that 'can take around 30 minutes' and warns about contradictory instruction following. Clarifai's infrastructure guide emphasizes that self-hosting requires 'eight A100/H100/H200 GPUs' and that 'even the INT4 version requires about 549 GB of VRAM.' Multiple sources recommend starting with API access and treating Agent Swarm as a 'capable junior teammate, not an autonomous operator.'
**Open-weight significance**: Several reviewers note this is 'the first time that the leading open weights model has supported image input,' removing a critical adoption barrier. The Modified MIT license with attribution requirements only above 100M MAU or $20M revenue is considered 'effectively unrestricted' for the vast majority of developers.
ユースケース
### 1. Multi-Document Research & Competitive Intelligence
Agent Swarm shines when tasks naturally decompose into parallel subtasks. Example: analyzing 20 competitor websites simultaneously, with each sub-agent handling one company. Testing confirmed ~4× speedup (14 min → 3.5 min). Choose K2.5 Instant over GPT-5.2 or Claude when the task is parallelizable, research-heavy, and cost-sensitive. The Agent Swarm mode has no equivalent in competing models.
### 2. Vision-to-Code Workflows (UI/Frontend Generation)
K2.5's native MoonViT encoder enables direct conversion of screenshots, Figma mockups, or hand-drawn wireframes into functional React/CSS code. One reviewer notes it 'reconstructed websites from video walkthroughs and converted hand-drawn sketches into functional 3D models.' Choose K2.5 Instant over alternatives when your pipeline involves visual inputs that need to become code — the end-to-end multimodal training produces more accurate layout and responsive breakpoints than adapter-based approaches.
### 3. High-Volume Document Processing & OCR
With 92.3% on OCRBench and 88.8% on OmniDocBench, K2.5 excels at extracting structured data from PDFs, invoices, and mixed-format documents. The 256K context window can ingest ~500 pages of text in a single prompt. At $0.60/1M input tokens (with cache hits at $0.10/1M), processing large document volumes is economically viable where Claude ($15/1M) or GPT-5.2 ($2.50/1M) would be prohibitive. Choose K2.5 Instant for batch document extraction where throughput and cost matter more than the absolute highest accuracy.
### 4. Real-Time Coding Assistant (Instant Mode)
The Instant variant is specifically designed for latency-sensitive coding workflows: quick code suggestions, refactoring, test generation, and component scaffolding. Its LM Arena Code Elo of 1526 places it in the top 10 coding models. The OpenAI-compatible API means drop-in integration with existing toolchains. Choose K2.5 Instant over Thinking mode (or Claude/GPT) when response speed is critical and the task doesn't require deep multi-step reasoning — typical savings of up to 75% fewer tokens compared to Thinking mode.
最新ニュース
• **Price Reduction (March 2026)**: API input pricing dropped from $0.60 to $0.45 per million tokens, as reported by OpenAIToolsHub, making it one of the cheapest frontier-class APIs available.
• **K2.5.1 Stability Update (March 2026)**: Moonshot released a minor update improving Agent Swarm stability for tasks spawning more than 50 parallel sub-agents, fixing a ~20% failure rate due to coordination timeouts.
• **Kimi Code CLI Launch**: A new open-source coding assistant integrating with VSCode, Cursor, and Zed editors was released alongside K2.5, available at kimi.com/code.
• **Deep Research Mode on kimi.com**: The web platform added a 'Deep Research' mode combining Thinking and Agent Swarm for multi-source fact-checking, though it consumes significantly more credits.
• **LM Arena Integration (February 2, 2026)**: Both Kimi-K2.5 and Kimi-K2.5-Instant were added to the Code Arena leaderboard, entering the top 10 within days of release.
• **WorldVQA Benchmark Release**: Moonshot open-sourced a new vision-centric world knowledge benchmark at github.com/MoonshotAI/WorldVQA, where K2.5 scores 46.3% (vs GPT-5.2: 28.0%, Gemini 3 Pro: 47.4%).