China's Moonshot AI has fired the loudest open-source salvo of the year. On July 16, the Beijing-based lab released Kimi K3, a 2.8-trillion-parameter Mixture-of-Experts model with native vision, a one-million-token context window and always-on reasoning — and pledged to publish the full weights on Hugging Face by July 27 under a modified MIT license. If the weights land as promised, K3 will be the largest open-source AI model ever released, dwarfing DeepSeek's 1.6-trillion and Alibaba's 397-billion-parameter offerings.
The First Open "3T-Class" Model
Moonshot is branding K3 as the world's first open 3T-class model, a reference to its parameter count crossing into a new scale tier. Under the hood, the model leans on two architectural innovations. Kimi Delta Attention is a hybrid linear attention mechanism designed to keep million-token contexts computationally tractable, while Attention Residuals act as a drop-in replacement for standard residual connections. Despite the enormous total parameter count, sparsity keeps inference manageable: just 16 of 896 experts activate per token.
For now, K3 is API-only, priced at $3 per million input tokens on cache misses, $0.30 per million cached input tokens and $15 per million output tokens. Until the Hugging Face release arrives, K3 is best described as open-weight in commitment rather than in practice — a distinction worth watching over the next week.
Benchmarks: Fourth in the World, First in Frontend
Independent evaluations broadly back Moonshot's frontier claims, with some nuance:
- Artificial Analysis scores K3 at 57.11 on its Intelligence Index, 76.24 on coding and 50.07 on agentic tasks — fourth out of 189 tracked models and on par with Claude Opus 4.8 and GPT-5.5.
- On LMArena's Frontend Code Arena, K3 seized the number-one spot with 1,679 points, edging out Claude Fable 5 (1,631), GPT-5.6 Sol (1,618) and GLM-5.2 (1,587) — a 17-place leap from its predecessor Kimi K2.6.
- Vals AI ranks K3 second overall behind Claude Fable 5, ahead of GPT-5.6 Sol, with a Vals Index of 74.7 and a strong 80.9 on Terminal-Bench 2.1.
Moonshot's own self-reported numbers show K3 beating Claude Opus 4.8 and GPT-5.5 on most tests while trailing the newest flagships, Claude Fable 5 and GPT-5.6 Sol. Cost-wise, Artificial Analysis estimates roughly $0.94 per task — close to GPT-5.6 Sol and about half of Claude Opus 4.8, though still well above smaller open-weight peers.
Timing Is Everything
The release landed strategically, just ahead of the 2026 World Artificial Intelligence Conference in Shanghai, and amid reports that Moonshot is raising between $1 billion and $2 billion at a valuation of up to $31.5 billion. It also arrives as the gap between open and closed models shrinks to its narrowest point ever: trackers now log a notable new model roughly every three days, and open-weight systems like DeepSeek V4 and GLM-5.2 already sit within striking distance of proprietary leaders on several coding benchmarks.
Bloomberg framed K3 as China's clearest signal yet that its labs are closing the distance with U.S. rivals — not through secrecy, but by giving frontier-scale technology away. Moonshot has momentum on the distribution front too: its earlier Kimi K2.7 Code model is already available directly inside GitHub Copilot, giving the lab a foothold in Western developer workflows that few Chinese competitors can match. A frontier-class K3, backed by fresh billions in funding, would extend that reach considerably.
Why It Matters
Kimi K3 resets the ceiling for what "open source" means in AI. Until now, openness generally implied accepting a meaningful capability discount versus the best closed models; K3 compresses that discount to a few benchmark points at the frontier — and eliminates it entirely in frontend coding. For enterprises, a self-hostable 3T-class model with a million-token context changes the build-versus-buy calculus, particularly in regulated industries wary of routing data through hosted APIs, whether American or Chinese.
The geopolitical subtext is just as significant. Weeks after U.S. export-control actions demonstrated that closed frontier models can be switched off by regulatory order, China's leading labs are pushing capability into artifacts that, once downloaded, cannot be recalled. That asymmetry — controllable APIs in the West, uncontrollable weights from the East — may prove the defining strategic tension of this AI cycle. The July 27 weights drop is now one of the most consequential dates on the AI calendar.
