Moonshot unveils Kimi K3, a 2.8-trillion-parameter open model that narrows the gap with US frontier AI

Chinese AI startup Moonshot on Friday unveiled Kimi K3, a 2.8-trillion-parameter model it describes as the world's largest open-weight AI system. The launch, reported by Reuters and others, places the Alibaba-backed firm directly alongside US frontier labs Anthropic and OpenAI, whose most advanced models it is now chasing.

What Kimi K3 brings

K3 is a sparse mixture-of-experts design that activates roughly 50 billion of its 2.8 trillion parameters per token by routing through 16 of 896 experts. It carries a 1-million-token context window and ships with a mechanism Moonshot calls Kimi Delta Attention, which the firm says decodes up to 6.3 times faster over million-token inputs. The architecture also includes a technique labelled Attention Residuals, credited with about 25% higher training efficiency at under 2% extra cost. Moonshot claims K3 achieves roughly 2.5 times better scaling efficiency than last year's Kimi K2.

Kimi K3 performed competitively with Fable 5 (with fallback) and substantially outperformed Opus 4.8, GPT 5.6 Sol, and GPT 5.5 in terms of GPU kernel optimisation.

The model will be fully open-sourced by the end of July 2026, with weights scheduled for release on 27 July. Until then, no outside researcher can confirm the parameter count or reproduce the benchmarks.

Benchmark scores and independent testing

On the company's own benchmarks, K3 ranks second overall behind only Anthropic's Fable 5 and OpenAI's GPT-5.6 Sol. Arena.ai ranked Kimi K3 first in a benchmark assessing web interface-building capabilities, while Vals AI placed it second overall behind Fable 5 and ahead of GPT-5.6 Sol. Artificial Analysis said the model delivered performance comparable to OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.8, particularly on tests measuring complex, multi-step tasks. Where independent testing does exist, K3 scored 77.8 on Program Bench.

The wider Chinese AI surge

Moonshot's launch is the latest in a series of moves by Chinese firms that are reshaping the global AI landscape. According to Fortune, Chinese open-source models led by Qwen, MiniMax and DeepSeek now account for one-third of global large language model usage, up from a near-zero presence at the end of 2024. DeepSeek recently closed a funding round of roughly $7 billion at a $52 billion valuation and is already negotiating a second round expected to push that figure to $71 billion, SCMP reported. Reuters confirmed that DeepSeek is developing its own ASIC chip for inference, aiming to reduce dependence on Nvidia and Huawei.

We are a group of very normal people.

Hong Kong-listed MiniMax is developing its own 2.7-trillion-parameter model for release as soon as the third quarter of 2026 and plans to launch its frontier-level multimodal model H3 in the near future. Z.ai, commonly known as Zhipu, is on track for annual recurring revenue of $1 billion, having achieved its full-year sales target in July.

Market reaction and US context

The launch comes a month after Anthropic's Fable and Mythos models were abruptly withdrawn by the US government due to security concerns. WirtschaftsWoche noted that the announcement is drawing parallels to the so-called "DeepSeek moment" of early 2025, when a Chinese model first shook assumptions of US dominance and sent global tech stocks tumbling. Bloomberg reported that the breakthrough rippled through global markets on Friday, sending AI-related shares lower.

What comes next

The weights for Kimi K3 are due on 27 July, after which independent researchers will be able to verify Moonshot's claims. The model's release is likely to intensify an already accelerating release cycle among Chinese AI firms, with MiniMax's 2.7-trillion-parameter system expected within months and DeepSeek's potential IPO on the horizon. Huawei is already integrating DeepSeek into its cloud services in sub-Saharan Africa, extending the reach of Chinese open models into new markets.