Chinese AI Startup Moonshot Unveils Largest Open AI Model

News

1. Chinese AI Startup Moonshot Unveils Largest Open AI Model

2. C-DAC Begins Trials of Indian AI Chip: Report

3. HCL Tech Buys Guardian India GCC in $10.5 Million Deal

4. Xi Calls for Global AI Cooperation

5. India Expands Chip Design Push Under Semicon 2.0

6. AI Dispatch | India Expands Chip Push With $13.2 Billion Plan

Chinese artificial intelligence startup Moonshot AI unveiled a 2.8 trillion-parameter model on Friday, July 17, stepping up competition with leading US developers and sending shares of two listed Chinese rivals sharply lower.

The Beijing company described Kimi K3 as the world’s first open model in the three-trillion-parameter class. It has a one-million-token context window and is designed for long coding sessions, knowledge work, reasoning and tasks involving text, images and video.

Kimi K3 is available through Moonshot’s website, desktop and coding products and its application programming interface. The company said it would release the model’s full weights by July 27, allowing developers to download, modify and deploy it on their own infrastructure.

Moonshot has not yet released the weights or a full technical report, however, limiting researchers’ ability to test its architectural and performance claims independently.

The company acknowledged that Kimi K3’s overall performance still trails Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol. It nevertheless said the model reached a level comparable with leading proprietary systems across several coding, reasoning and agent-based evaluations.

In Moonshot’s tests of GPU-kernel optimization, Kimi K3 performed competitively with Claude Fable 5 and substantially outperformed Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.6 Sol and GPT-5.5.

GPU-kernel optimization measures a model’s ability to rewrite low-level software so that it uses computing hardware more efficiently and reduces latency. Moonshot said the models were tested in identical environments and given as long as 24 hours to complete four assignments.

Those results were produced or compiled by Moonshot and should not be treated as fully independent comparisons. The company also noted that Claude Fable 5 was evaluated by a third party and that its results may have included fallback to another model.

Separate evaluations indicated that Kimi K3 performed strongly on some tasks. Reuters reported that Arena.ai ranked it first in a test of web-interface development, while Vals AI placed it second overall behind Claude Fable 5 and ahead of GPT-5.6 Sol. Artificial Analysis found its performance comparable with GPT-5.5 and Claude Opus 4.8 on complex, multistep work.

Benchmark results can vary according to test conditions, software harnesses, prompts and reasoning settings. Moonshot ran Kimi K3 at its maximum reasoning setting, and some comparisons used different coding systems for different models.

Kimi K3 uses a mixture-of-experts architecture containing 896 experts, of which 16 are activated for a task. Such systems contain large numbers of parameters but use only a portion of them for each operation, reducing the computing demands compared with activating the full model.

Moonshot said Kimi Delta Attention, Attention Residuals and changes to its training methods gave Kimi K3 approximately 2.5 times the scaling efficiency of its earlier K2 model. The company defines that as an improvement in its ability to convert additional computing power into model capability.

The model’s size may still limit the practical value of its open release. Running a 2.8 trillion-parameter system privately would require extensive computing infrastructure, putting self-hosting beyond the reach of most individual developers and smaller companies.

Moonshot recommends deploying Kimi K3 on systems containing at least 64 accelerators. Ryan Fedasiuk, a fellow at the American Enterprise Institute, estimated that the equipment required to operate it locally could cost hundreds of thousands of dollars, Reuters reported.

Moonshot charges $3 per million uncached input tokens and $15 per million output tokens through its API. Cached inputs cost 30 cents per million tokens. The company says more than 90% of inputs are served from its cache in coding workloads, although that rate may not apply to every customer or use case.

The announcement weighed heavily on Moonshot’s listed Chinese competitors. Zhipu AI, which uses the Z.ai brand, was down 27.7% in Hong Kong shortly before the close, while MiniMax fell 16.5%, according to Reuters.

The falls suggested that investors were concerned Kimi K3 could weaken competitors’ technological position and intensify price competition in China’s crowded AI market.

Technology and semiconductor stocks also fell across global markets on Friday, but Moonshot’s announcement was only one part of a broader retreat. The selloff reflected doubts about returns on AI infrastructure spending, weakness in chip shares, expectations of higher US interest rates and rising oil prices amid the US-Iran conflict.

Japan’s Nikkei 225 closed 4.03% lower, entering correction territory, while the Philadelphia Semiconductor Index had fallen 4.3% in the preceding US session. US stock-index futures also pointed lower.

Before Kimi K3, China’s largest disclosed models included Meituan’s LongCat 2.0 and DeepSeek’s V4-Pro, each with about 1.6 trillion parameters.

Parameter counts measure a model’s scale rather than its intelligence or practical usefulness. Comparisons with US systems are also difficult because OpenAI and Anthropic do not disclose the parameter counts of their latest models.

Founded in 2023 by AI researcher Yang Zhilin, Moonshot is backed by investors including Alibaba and Tencent. Bloomberg reported in June that the company was seeking as much as $2 billion at a valuation of about $30 billion before a possible Hong Kong listing.

Kimi K3 adds to evidence that Chinese developers are narrowing the performance gap with US laboratories despite Washington’s restrictions on China’s access to advanced semiconductors and chipmaking equipment.