Moonshot AI Unveils Kimi K3, an Open 2.8T-Parameter Frontier Model
On July 16, 2026, Beijing’s Moonshot AI launched Kimi K3—a 2.8-trillion-parameter open-class multimodal model with a 1M-token context window, API live the same day, and full weights planned for July 27—claiming frontier coding and agent scores just behind Claude Fable 5 and GPT-5.6 Sol at far lower token prices.
TLDR
Moonshot AI on July 16, 2026 introduced Kimi K3, described as the world’s first open 3T-class model: 2.8 trillion parameters, native vision, 1-million-token context, built on Kimi Delta Attention (KDA) and Attention Residuals. The company says overall performance still trails Claude Fable 5 and GPT-5.6 Sol but beats other tested models across coding, agentic, and multimodal suites. API is live on the Kimi platform (kimi-k3); full weights target July 27, 2026. Official API pricing: $0.30 / $3.00 / $15.00 per MTok for cache-hit input, cache-miss input, and output.
Specs and availability
From Moonshot’s primary tech blog (kimi.com/blog/kimi-k3) and product surfaces:
| Item | Detail |
|---|---|
| Parameters | 2.8T (open 3T-class); Stable LatentMoE activates 16 of 896 experts |
| Context | 1M tokens |
| Architecture | KDA + AttnRes; ~2.5× scaling efficiency vs Kimi K2 (company claim) |
| Quantization | MXFP4 weights / MXFP8 activations from SFT onward |
| Surfaces | kimi.com, Kimi Work, Kimi Code, Kimi API; app stores (iOS/Android/HarmonyOS) |
| Default effort | Max thinking at launch; low/high modes later |
| Weights | Full release by July 27, 2026 |
| Deploy note | Supernode guidance: 64+ accelerators recommended |
Selected company-reported max-effort benchmarks place K3 near Fable 5 / Sol on coding (e.g. Program Bench 77.8, Terminal Bench 2.1 88.3, SWE Marathon 42.0) and agentic browse/search tasks, with honest footnotes that harnesses differ by model.
Independent rankings (Arena, Artificial Analysis)
Third-party boards—not Moonshot’s marketing table—are where K3 landed hardest:
| Source | Snapshot | Standing |
|---|---|---|
| LMArena Frontend Code Arena | Unblinded launch window (Jul 16) | #1 at 1679 Elo — ahead of Claude Fable 5 (moonshot-kimi-k3-tops-arena-frontend-code |
| Arena Text Overall | Board dated Jul 16, 2026 (arena.ai) | #9 at 1486 ±11 (preliminary, ~3,024 votes) — tied band with GPT‑5.6 Sol xhigh; Claude Fable 5 still #1 at 1507 |
| Artificial Analysis Intelligence Index | AA post Jul 17, 2026 | 57 — #3 overall; comparable to Opus 4.8 / GPT‑5.5; still behind Fable 5 and GPT‑5.6 Sol ( |
Labeling: vendor benches are harness-specific; Arena scores are blind human preference; AA is a paid eval suite. Open weights (target Jul 27) would make K3 the clear open-weight Intelligence Index leader among released peers (GLM-5.2 51, DeepSeek V4 Pro 44 per AA)—until then Arena lists it as proprietary API access.
Price and market shock
Fortune and Bloomberg same-week coverage framed K3 as a potential “second DeepSeek moment”: open (or soon-open) weights, Fable-competitive coding claims, and output priced at $15/MTok versus reported ~$50 for Fable and ~$30 for GPT-5.6 Sol—still above cheap Chinese peers (GLM-5.2, DeepSeek V4) but dramatically under U.S. closed flagships. Moonshot raised a $2B round in May at >$20B valuation (Meituan-led; Alibaba/Tencent/HSG among backers); ARR was cited above $200M in advisor statements. Prior Kimi versions already appear in Western stacks (Cursor Composer 2; DoorDash lower-level work; Thinking Machines used Kimi K2.5 for early Inkling post-train data per Fortune).
Why this story matters
K3 compresses the U.S.–China frontier gap on an open path just as WAIC opens in Shanghai and U.S. export-control debates intensify. If independent evals hold the company table, open Chinese models are no longer “good enough and cheap”—they are near-frontier and still cheaper. That pressures closed-lab pricing, complicates distillation accusations, and forces enterprises to treat Chinese open weights as a real procurement option—not a science experiment.
Sources
- Moonshot AI: “Kimi K3: Open Frontier Intelligence” tech blog (July 16, 2026) — kimi.com/blog/kimi-k3
- Arena.ai / LMArena: Frontend Code Arena #1 (1679 Elo) posts Jul 16; Text Overall leaderboard snapshot Jul 16, 2026
- Artificial Analysis: “Kimi K3 achieves #3 in the Artificial Analysis Intelligence Index…” (July 17, 2026)
- OpenLM.ai model card summary and benchmark table (dated to July 16, 2026 evaluation snapshot)
- Fortune: “Moonshot’s Kimi K3 pushes Chinese AI into Fable-level territory” (July 16, 2026); market-follow Bloomberg/Fortune pieces (July 17, 2026)