Moonshot AI has released the full weights of Kimi K3 on Hugging Face today, July 27, 2026 — delivering on a promise made when the model debuted at the World AI Conference in Shanghai on July 16. At 2.8 trillion parameters, K3 is the largest open-weight model ever published, and its arrival reshapes the economics and geopolitics of AI deployment for enterprises and developers worldwide.
Key Highlights
- 2.8 trillion parameters — the largest open-weight model release in history, surpassing all previous open-source milestones
- 16 of 896 experts fire per token, keeping active compute at roughly 50 billion parameters for efficient inference
- 1-million-token context window with native vision capabilities baked in
- Modified MIT license — organizations can download, inspect, fine-tune, and self-host
- #2 on the Vals AI Index and #1 in Frontend Code Arena among all publicly evaluated models
- Performance gap with frontier closed from 6–9 months to roughly 3–5 months
Architecture: KDA and Stable LatentMoE
K3's architecture combines Kimi Delta Attention (KDA) and Attention Residuals (AttnRes) with a Stable LatentMoE design — Moonshot's second-generation mixture-of-experts framework. Only 16 of 896 experts activate per token, making the model sparse and theoretically efficient despite its enormous parameter count. Moonshot reports a 2.5× improvement in overall scaling efficiency compared to Kimi K2.
Native vision is built into the base model rather than bolted on through an adapter, enabling K3 to reason over images and documents within the same context as code and text — a capability that rivals proprietary multimodal leaders.
Self-Hosting: Powerful but Demanding
Downloading K3 is free. Running it is another matter. The MXFP4 four-bit precision weights alone require approximately 1.4 terabytes of fast GPU memory before runtime buffers and activations are factored in. Moonshot recommends production deployments use "a supernode configuration with 64 or more accelerators" — placing real-world K3 inference firmly in the data-center tier, beyond the reach of individual workstations or even well-equipped on-premise server rooms.
For organizations that can clear that infrastructure bar, self-hosting removes prompts from Moonshot's API entirely, eliminating the data sovereignty concern that many regulated industries associate with Chinese-operated AI APIs under the National Intelligence Law.
Benchmarks: Near the Frontier
K3's numbers on public leaderboards are striking:
- Artificial Analysis Intelligence Index: 57, ranked 4th of 189 evaluated models
- Vals AI Index: 74.70%, ranked 2nd of 38
- Arena WebDev: 1st place with a preliminary score of 1,679
- Output speed: 62 tokens per second via hosted API
- First-token latency: 1.99 seconds
The model consistently outperforms all previous open-weight releases on coding benchmarks — DeepSWE, Terminal-Bench, and Program Bench — while trailing the top proprietary models (Claude Fable 5 and GPT-5.6 Sol) on overall reasoning evals. Moonshot's blog describes one research scenario where K3 "completed in about two hours what would typically require one to two weeks" of expert work.
API Pricing vs. Self-Hosting Cost
Organizations choosing to use Moonshot's hosted API pay:
| Token type | Cost |
|---|---|
| Cache-hit input | $0.30 per million tokens |
| Cache-miss input | $3.00 per million tokens |
| Output | $15.00 per million tokens |
That's roughly three to four times more expensive than Kimi K2.6, reflecting K3's expanded capabilities and context length. Self-hosted inference eliminates per-token fees but replaces them with capital expenditure on a 64-accelerator (or larger) cluster — a trade-off that favors high-volume enterprise workloads.
The Open-Weights Geopolitical Signal
Analysts observing the release note that it arrives alongside Chinese President Xi Jinping's public commitment to open-source AI and global capability diffusion. By releasing K3's weights under a permissive license, Moonshot enables any researcher or company to study, fine-tune, and redistribute the model — creating an asymmetric access dynamic where powerful Chinese-origin open models proliferate globally while comparable U.S. models remain closed or export-restricted.
The interconnects.ai analysis characterizes K3 as evidence that Chinese labs have achieved "near-frontier" capabilities with "far, far fewer resources" than Western counterparts, accelerating the global diffusion of advanced AI in ways that make export controls difficult to enforce.
What's Next
The immediate community focus will be on:
- vLLM integration — Moonshot is expected to publish a vLLM implementation with native Kimi Delta Attention support alongside the weights, enabling standard inference tooling
- Fine-tuned variants — the open license invites the research community to produce domain-specific checkpoints
- Quantized serving — inference providers are already working on INT4 and INT8 variants to bring per-token costs closer to GPT-4o parity
- Regulated-industry audits — the ability to inspect weights addresses model transparency requirements in healthcare, finance, and government procurement frameworks
Source: Moonshot AI — Kimi K3 Blog