On 17 July 2026, on stage at the World Artificial Intelligence Conference in Shanghai, Moonshot AI unveiled Kimi K3: the first open-source model to reach 2.8 trillion parameters. A number that until a few months ago seemed the exclusive preserve of the major American laboratories.
Financial markets reacted accordingly: TSMC dropped 7%, Nvidia lost 1.2%, the Nasdaq 100 fell a full percentage point. The second “DeepSeek shock” in less than two years.
2.8 trillion parameters, open source
Kimi K3 is built on two architectural innovations developed in-house by Moonshot: Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), which improve how information flows across sequence length and model depth. The system uses a high-sparsity Mixture of Experts (MoE) approach: 16 active experts out of 896, managed through a Stable LatentMoE framework.
The result is a scaling efficiency roughly 2.5 times better than Kimi K2. The model supports a 1-million-token context window and has native multimodal capabilities.
The full weights will be released on 27 July 2026, making Kimi K3 the largest open-source model ever published.
Where it stands against frontier models
Moonshot is transparent about the limitations: K3 does not surpass Claude Fable 5 and GPT-5.6 Sol in overall evaluations. But it approaches them credibly, and in some specific areas it surpasses both.
On long-horizon coding benchmarks - those measuring real engineering sessions, not just code snippets - K3 shows results competitive with Fable 5 and substantially outperforms Opus 4.8 and GPT-5.6 Sol. On BrowseComp, the agentic web navigation benchmark, it reaches 90.4% with the 1-million-token window.
Particularly impressive is the practical demonstration: during its own development, an early version of K3 autonomously handled the majority of the team’s GPU kernel optimisation work. In another demo, it built MiniTriton from scratch - a Triton-compatible compiler with its own PTX pipeline, delivering performance comparable to or better than Triton on some workloads.
Arena.ai, in blind tests using human preferences, ranked K3 first for web interface engineering, ahead of Fable 5.
The price: one third of Fable 5
This is the point that shook the markets. Kimi K3 is available via API at:
- $0.30/MTok input (cache hit)
- $3.00/MTok input (cache miss)
- $15.00/MTok output
Claude Fable 5 costs $50/MTok on output. A gap of more than three times, with performance that comes remarkably close. It is not hard to understand why DoorDash, Cursor and other American players are already integrating Kimi models into their products.
A model designed to work autonomously
Kimi K3 was built to operate with minimal human supervision on extended tasks: engineering sessions on massive codebases, tool orchestration, academic research with dozens of parallel subagents. Moonshot explicitly describes “long-horizon coding” and “knowledge work” as its primary use cases.
The company also acknowledges a limitation: K3 tends to be excessively proactive, making decisions autonomously even when the user would prefer to be involved. For applications requiring constrained behaviour, they recommend explicit constraints in the system prompt or AGENTS.md.
What it means for the market
K3 is yet another confirmation of a pattern that is consolidating: Chinese laboratories release frontier open-source models at structurally lower prices than American competitors. The reason is a combination of lower energy costs, margins consciously sacrificed to capture market share, and years of work to extract maximum performance from less powerful hardware due to US export blocks.
President Xi Jinping, speaking at the same WAIC, reaffirmed China’s commitment to open source in AI - with a direct dig at Washington’s export control policies.
Kimi K3 on AIDeskPro
Kimi K3 is available in the Kimi app and via API from today. The model runs on global endpoints (China/US) and in AIDeskPro it is labelled “No EU”: it can be freely chosen for non-critical operations on non-sensitive information, but it is not among the Tool’s default models. As always, the choice is yours and always informed.
Source: Kimi K3: Open Frontier Intelligence - Moonshot AI , 17 July 2026.