Moonshot AI (Kimi)
Moonshot AI (Kimi)
One-line summary: Chinese frontier lab behind the Kimi model series; its Kimi K3 (released 2026-07-18, weights ~2026-07-27) hit #1 open-weight and #3 on the overall cost/performance frontier, triggering an "AI Sputnik moment."
What it is
Moonshot AI is a Chinese lab (CEO Yang Zhilin, a Carnegie Mellon PhD) whose Kimi series (K2 → K2.6 → K2.7 → K3) climbed the open-weight leaderboards over ~a year. Per the panel in 2026-07-19-podcast-moonshots-urgent-update-ai-sputnik-moment-kimi-k3-released, the lab claims Kimi models held state-of-the-art among open-weight models in 9 of the past 12 months. Company valuation ~$20B (vs. ~$1T for OpenAI/Anthropic).
Kimi K3 specs (from the pod): 2.8T total parameters / ~50B active (MoE), multimodal, #1 on the front-end code arena (past Claude Fable 5) and #1 in six other domains; #3 on the Artificial Analysis Intelligence Index cost/performance Pareto frontier. Architecturally "no magic — still essentially a transformer" (alexander-wissner-gross) with a Muon-optimizer + data-mix edge; trained on H800s but optimized to run on Huawei Ascend / Alibaba chips. Full weights slated for ~2026-07-27.
Why it matters to artificial-intelligence
Kimi K3 is the strongest evidence yet for chinese-open-weight-frontier-parity: a Chinese open-weight model reaching the overall frontier (not just a benchmark spike), at <½ the token cost, engineered around US chip export controls. emad-mostaque: "building great solid models is cutting edge manufacturing... why are Chinese EVs better than Ford's? This actually feels like the same thing." It also underwrites frontier-intelligence-perishable (open weights track within a release) and the edge-inference-shift (distill-to-edge). Xi Jinping's WAIC speech (same week) framed open source as a public good China will "fully back," with model-approval time cut from ~60 days to ~1 week.
Why it matters to stock-market
Moonshot's ~$20B valuation vs. the ~$1T frontier labs is the crux of the derate trade: dave-blundin cites gavin-baker that "all businesses, all stocks other than the foundation AI labs are huge beneficiaries," while foundation-lab valuations get re-questioned and silicon demand rises (Jevons). Canonical chain: open-weight-sputnik-to-frontier-lab-derate; margin-transfer logic: open-source-share-shift-bullish-for-compute.
Key facts
- Kimi K3: 2.8T params / ~50B active, multimodal, #1 open-weight + #3 overall Pareto frontier; weights ~2026-07-27.
- Kimi K3 API ~$15/M tokens (vs. DeepSeek ~$1, Opus ~$40, Fable ~$60) — high because run on less-efficient Chinese silicon at high margin; expected to drop 10-50x as US inference providers optimize it.
- CEO Yang Zhilin: CMU PhD (2019); prior startup Recurrent AI (China).
- Valuation ~$20B.
Adjacent Code Arena color (2026-09-02) — not a K3 rewrite
- From 2026-09-03-x-ai-pass-3-sep-2026-open-model-fleets-agentic-coding-price (@arena): Code Arena WebDev lists Kimi K3 (Max) 17 pts behind qwen-3-8-max-0902 (1,691) and 14 pts behind Claude Opus 5 (Max) at that snapshot. Chronological color on a later board; do not rewrite the July K3 "front-end code arena #1" podcast specs above. Different board, different date, Max SKU as written by the arena operator.
Adjacent AA Index v4.3 color (2026-09-07) — not a July K3 rewrite
- From 2026-09-07-artificial-analysis-intelligence-index-v4-3 (@ArtificialAnlys): on Intelligence Index v4.3, Kimi K3 44 (tied with GLM-5.3 as open-weights co-lead); 9 points behind gpt-6-astra / claude-fable-5-1 at 53. Different Index version than the July podcast “#3 on the overall cost/performance frontier.” Do not rewrite the July specs above. Full treatment: artificial-analysis-intelligence-index.
Adjacent Anthropic distillation allegation (2026-09-10) — issuer assessment, not a K3 rewrite
- From 2026-09-11-anthropic-sep-2026-threat-intelligence-autonomous-cyber (issuer threat report): Anthropic attributes GTG-16002 to Moonshot / Kimi — silent relay of customer requests to Claude displayed as Kimi; May–Jul >23 million exchanges; one 10-day window ~300,000 requests via 5,380 fraudulent accounts. Sensitive-customer / PLA-affiliated examples are Anthropic investigative narrative. Named-lab reply not retrieved. Do not rewrite July K3 specs above. Full treatment: anthropic-illicit-distillation-sep-2026.
Open questions
- Is K3 "benchmaxed"? (Panel: open weights will settle it within ~2 weeks of release.)
- Will the US move to block Kimi weights from Hugging Face / restrict corporate use? See ai-self-regulatory-body.
Sources
- 2026-07-19-podcast-moonshots-urgent-update-ai-sputnik-moment-kimi-k3-released (multi-context, vault/sources/)
- 2026-09-03-x-ai-pass-3-sep-2026-open-model-fleets-agentic-coding-price — Code Arena WebDev color only (Kimi K3 Max, 17 pts back)
- 2026-09-07-artificial-analysis-intelligence-index-v4-3 — Index v4.3 Kimi K3 44; not a July rewrite
- 2026-09-11-anthropic-sep-2026-threat-intelligence-autonomous-cyber — Anthropic distillation allegation (issuer); not a K3 rewrite
Related
- chinese-open-weight-frontier-parity
- open-weight-sputnik-to-frontier-lab-derate
- open-source-share-shift-bullish-for-compute
- frontier-intelligence-perishable
- emad-mostaque
- thinking-machines-lab
- qwen-3-8-max-0902
- artificial-analysis-intelligence-index
- glm-5-3-flash
- anthropic-illicit-distillation-sep-2026
- anthropic-sep-2026-threat-intelligence