jimmy·research
Home/AI Infrastructure/Kimi
EN↔中文
Product

Kimi

Moonshot AI's flagship long-context LLM series; one of the leaders in China's 2026 open-source agentic coding models

1. Core Product / Service

Kimi is Moonshot AI's core LLM product line. Since its 2023 launch, long context has been the differentiated selling point (early 200k tokens, far above peers at the time).

Current flagship version (2026-05):

  • Kimi K2.6 — flagship open-source model released 2026-04-20, 1T-parameter MoE architecture, 32B activated per token, native multimodal + agentic capability, 256K context, max output 16K tokens [openrouter.ai/moonshotai/kimi-k2.6, 2026-05-09]
  • Kimi K2.5 Turbo / K2.5 Thinking — previous generation (still in product line); Turbo for fast response, Thinking for reasoning depth
  • Kimi K2 Turbo — older flagship, still has users

Key capabilities:

  • Agent Swarm system: can dispatch up to 300 domain sub-agents, autonomously executing 4000 steps in a single run
  • End-to-end coding (Python / Rust / Go) + UI/UX generation
  • Benchmarks: SWE-Bench Verified 80.2, Terminal-Bench 2.0 66.7, SWE-Bench Pro 58.6 (beating GPT-5.4 xhigh 57.7, Claude Opus 4.6 max 53.4)

Deployment forms: official platform.moonshot.ai direct / OpenRouter / DeepInfra / Cloudflare Workers AI / vLLM + SGLang self-deployment.

Official pricing (K2.5 Turbo, reference): input $0.60/M (cache hit $0.15/M), output $2.50/M; K2 Turbo input $1.15/M, output $8.00/M.

2. Target Users & Pain Points

Target users:

  • Chinese and international developers needing long context (processing entire books, entire code repos)
  • Practitioners of agentic coding workflows
  • High-frequency users not wanting to pay USD prices for Claude/GPT (China official pricing + third-party OpenRouter pricing significantly below US frontier models)

Pain points solved:

  • Long-document processing (256K context fits a mid-sized code repo)
  • Long-running autonomous agent tasks (K2.6's strength is long-horizon agentic rather than ambiguous one-shot reasoning)
  • Chinese developers going global bypassing international credit card payments via USDC and similar (see the openrouter integration discussion in hermes-openrouter-models)

3. Competitive Landscape

Model Camp Strengths Weaknesses
Kimi K2.6 Moonshot Long-horizon agentic, 1T MoE open source, SWE-Bench Pro leader Open-ended ambiguous planning is weaker
deepseek V4 Pro High-Flyer BenchLM overall #1 (87 points), price/performance Agent swarm not as strong as K2.6
Qwen 3.6 Plus Alibaba The only 1M context, Terminal-Bench leader Single-step reasoning slightly weaker than K2.6
GLM 5.1 Zhipu Strongest at front-end agentic work Overall slightly lower
Claude Opus 4.6 Anthropic Strongest overall High price, SWE-Bench Pro overtaken by K2.6
GPT-5.4 / 5.5 OpenAI General capability SWE-Bench Pro overtaken by K2.6

Kimi's differentiation: within China's open-source camp, K2.6 is the "long-distance runner" specialist; DeepSeek is the overall champion; Qwen pushes context limits; GLM specializes in front-end. These four are the first tier of Chinese open-source LLMs in 2026.

4. Unique Observations

  • Jimmy's Hermes integration: Moonshot Kimi connects to Hermes without source changes (hermes auth add kimi-coding); K2.6 is already in the built-in model list. See hermes-openrouter-models.
  • OpenRouter is cheaper than official: repeatedly observed that Kimi pricing on OpenRouter / DeepInfra third parties is lower than Moonshot official, making it the preferred price/performance path.
  • OpenClaw measured response time: Kimi K2.5 (Moonshot direct) averages 1.7-1.8s, on par with MiniMax M2.7 (OpenRouter), slower than Gemini 3.1 Pro direct (1.1s).
  • Historical fallback chain instability: while troubleshooting OpenClaw cron on 2026-04-11, found the kimi-k2.5 fallback chain failing during certain time windows; switched back to gemini-3-pro-preview.
  • Payment scenario judgment: in ai-inference-engines research, Kimi API was listed as a potential first-wave China-side adapter for Tempo / x402-style AI Agent-native payment protocols.

5. Financials (parent company Moonshot AI)

Date Event Valuation
2023-03 Founded (Yang Zhilin, Zhou Xinyu, Wu Yuxin, Tsinghua alumni) —
2024-02 Alibaba led $1B [techcrunch, 2026-05-09] $2.5B
2024-08 Tencent, Gaorong $300M $3.3B
2025-10 IDG led ~$600M (Tencent followed) $3.8B pre
2025-12 IDG $150M (Alibaba, Tencent followed) $4.3B
Early 2026 $700M $10B
2026-05-07 $2B round (open-source AI demand surge) [techcrunch, 2026-05-09] $20B post

Major investors: Alibaba (largest), Tencent, IDG Capital, Meituan, China Mobile, Gaorong Capital.

IPO rumors: Moonshot is reportedly considering a Hong Kong IPO, targeting a new round of $1B funding.

6. People & Relationships

  • Parent company / creator: Moonshot AI (月之暗面) — Kimi is its only flagship product line; the company and product are deeply bound at the entity level
  • Founder / CEO: Yang Zhilin (杨植麟, Tsinghua alumnus, ML scholar background)
  • Co-founders: Zhou Xinyu, Wu Yuxin
  • Major investors: Alibaba, Tencent, IDG Capital, Meituan, China Mobile
  • Competitors: deepseek (same tier of Chinese open source), Qwen (same investors as Alibaba but product competition), GLM (Zhipu), Claude, GPT
  • Distribution partners: openrouter (key third-party distribution channel), DeepInfra, Cloudflare Workers AI

Sources

Related

deepseekopenrouter
Last compiled: 2026-05-09