- Lead. China’s Moonshot AI released Kimi K3 on July 16, a 2.8-trillion-parameter open-weight model that ranks third on the GDPval-AA v2 benchmark — behind only Claude Fable 5 Max and GPT-5.6 Sol Max — making it the largest open-weight AI model yet released and placing a Chinese lab at the top tier of the global performance table.
- Fact. Priced at $3 per million input tokens and $15 per million output tokens, with full model weights scheduled for public release on July 27, Kimi K3 gives developers worldwide access to frontier-tier AI on private infrastructure at a fraction of the cost of comparable proprietary systems.
- Stake. The announcement triggered semiconductor stock declines, reprising the market reaction to DeepSeek’s model releases, as investors again questioned whether AI infrastructure spending at the scale currently underway in the United States is justified if Chinese labs can reach the same performance plateau on lower hardware budgets.
Moonshot AI, founded by Yang Zhilin, began rolling out Kimi K3 through the Kimi Code platform and consumer app on July 16. The model carries 2.8 trillion total parameters — the highest of any open-weight system released to date — and supports a context window of one million tokens, long enough to process an entire legal archive or a large software codebase in a single pass. On GDPval-AA v2, a benchmark measuring performance across 44 occupations and 9 industry categories, Kimi K3 scored 1,687, placing it third overall behind Claude Fable 5 Max at 1,815 and GPT-5.6 Sol Max at 1,747.8, according to CryptoBriefing.
Architecture and Efficiency
Moonshot highlighted two internal architectural innovations: Kimi Delta Attention, a hybrid linear attention mechanism designed to reduce compute requirements at extended context lengths, and Attention Residuals, a drop-in replacement for standard residual connections that delivers consistent gains as models scale. The company’s approach echoes the efficiency-first philosophy that Chinese AI developers have pursued since DeepSeek demonstrated that constrained hardware budgets need not translate into lower benchmark performance.
Kimi K3 is being released under a Modified MIT licence. Earlier Kimi versions had already been adopted by US enterprise users including Cursor, the AI coding environment, and DoorDash — signalling that enterprise developers were prepared to integrate Chinese open-weight models into production pipelines even as chip export restrictions continued to limit the hardware available inside China for training future generations.
The Market Reaction
Semiconductor stocks fell in the trading sessions immediately following the announcement, continuing a pattern that has accompanied Chinese AI model releases since DeepSeek’s first major benchmark victories. The iShares Semiconductor ETF pulled back more than 10% over the week ending July 17, as investors priced in the possibility that demand for high-end AI chips would moderate if competitive frontier models could be built at lower hardware cost.
China’s GLM-5.2, released by Zhipu AI earlier this year, had already beaten GPT-5.5 on coding benchmarks at a fraction of the cost. Kimi K3 extends that dynamic further: it is not merely competitive on a cost-efficiency curve but places in the top three of the overall performance leaderboard — a position that Western labs had treated as exclusively theirs until this year.
What the Open Release Means
Full model weights are scheduled for public release on July 27, giving enterprise developers and research institutions roughly ten days to prepare integration environments. Once released, Kimi K3 will be freely deployable on any private infrastructure globally. For AI governance bodies, the release of a 2.8-trillion-parameter model as open weights raises questions that existing frameworks were not designed to address: the EU AI Act’s GPAI provisions, which took effect in August 2025, were calibrated primarily for models in the hundreds-of-billions-of-parameter range, and no equivalent provision exists for models ten times that size deployed outside the EU by a non-EU developer. The July 27 release date will test how quickly those frameworks adapt.