Moonshot AI (Kimi)
Top signals
Agent answer
Moonshot AI (Kimi) has 246 loaded public signals: 7 hiring, 0 forks, 200 releases or model cards, 0 talking, and 39 repos. Latest signal: MoonshotAI/kimi-code @moonshot-ai/kimi-code@0.29.1. Data-business radar maps 2 signals to Evals and quality, Infrastructure. The standing analysis was generated with an unknown model and 0 evidence refs.
has loaded 246 public signals
has hiring signal count 7
has fork signal count 0
has release signal count 200
Thesis
Moonshot AI (Kimi) is shipping open-weight, trillion-parameter mixture-of-experts frontier models at a fast iteration cadence — the Kimi-K2 line is its flagship, now through K2.5 and K2.6 plus a dedicated K2-Thinking variant. Alongside the weights it is building a full agentic-coding surface (the kimi-cli / kimi-code tools) and publishing efficiency-oriented architecture research (linear attention, attention residuals, block attention). It is a Beijing-based lab pairing aggressive open releases with developer tooling.
Shipping
- Kimi-K2 family (≈1T-param MoE). The most-downloaded model is `moonshotai/Kimi-K2.6` at 3,139,192 30-day downloads (1,417 likes), followed by `Kimi-K2-Instruct-0905` at 2,734,600 and `Kimi-K2.5` at 1,682,758 (the single most-liked card at 2,813 likes). The original `Kimi-K2-Instruct` adds 629,908 downloads, a `Kimi-K2-Base` (42,722) is published, and a reasoning-focused `Kimi-K2-Thinking` has 163,935 downloads (1,699 likes). The Kimi-K2 GitHub repo is the lab's top repo at 10,839 stars.
- Multimodal (Kimi-VL). `Kimi-VL-A3B-Instruct` (297,546 downloads) and `Kimi-VL-A3B-Thinking` (134,822), with a `Kimi-VL-A3B-Thinking-2506` refresh and the `MoonViT-SO-400M` vision encoder; the Kimi-VL repo has 1,198 stars.
- Efficiency / architecture lines. `Moonlight-16B-A3B-Instruct` (101,711 downloads) and `Kimi-Linear-48B-A3B-Instruct` (61,936) — sparse A3B (≈3B active) designs.
- Audio & coding. `Kimi-Audio-7B-Instruct` (89,403 downloads; Kimi-Audio repo at 4,645 stars) and the SWE-oriented `Kimi-Dev-72B`.
- Developer tooling, actively versioned. `kimi-cli` (8,916 stars) is shipping rapidly — releases 1.44.0 through 1.47.0 — alongside the `kimi-code` package (2,041 stars), versioned 0.8.0 → 0.11.0, and a `walle` v0.1.10 release.
Research themes
No first-party writing captured yet. The themes are inferable only from open-source repos: efficiency-oriented attention/architecture research — Attention-Residuals (3,299 stars), MoBA (block attention, 2,123 stars), Kimi-Linear (linear attention, 1,399 stars), and the Kimi-k1.5 reasoning work (3,472 stars).
Hiring & scaling
All seven open roles are based in Beijing (北京), signaling a single-hub build-out. The mix is research-and-platform heavy: an algorithm researcher (算法研究员) and algorithm engineer (算法工程师) point to continued model R&D, while backend (后端开发工程师), frontend (前端开发工程师), product manager (产品经理), and designer (设计师) roles indicate investment in a productized, user-facing surface (consistent with the Kimi app and CLI tooling). One operations specialist (运营专员) role rounds out a go-to-market push rather than a pure research lab posture.
Traction highlights
- Hacker News: the lab's architecture research drew the most attention — Attention-Residuals at 241 points / 34 comments and Kimi-Linear at 217 points / 47 comments; smaller threads appeared for kimi-cli (5 pts), checkpoint-engine (2 pts), and FlashKDA (2 pts).
- Most-starred repos: Kimi-K2 (10,839), kimi-cli (8,916), Kimi-Audio (4,645).
- Most-downloaded models (30-day): Kimi-K2.6 (3.14M), Kimi-K2-Instruct-0905 (2.73M), Kimi-K2.5 (1.68M).
Three flagship cards, no leaderboard
A cross-card read of GPT-5.6, GLM-5.2 and Kimi-K2-Thinking: the closed leader publishes zero capability comparisons while the two open labs publish incompatible ones — so the leaderboard everyone wants cannot be built from primary sources. That void is the eval product.
Read the analysis →Deep reportKimi-K2-Thinking — the open agentic-research bet
Moonshot’s open 1T-MoE thinking agent holds the top of agentic web research (BrowseComp SOTA, 200–300 tool calls without drift) while trailing closed frontier on raw coding — a distinct open frontier from GLM-5.2’s agentic engineering.
Read the analysis →Data-business radar
cross-lab →2 matches · 2 active lanes
Moonshot AI (Kimi) has a repo signal matching infrastructure.