Moonshot says Kimi K3 will ship as a 2.8T-parameter open-weight model

Moonshot AI announced Kimi K3, a 2.8 trillion-parameter large language model whose weights are scheduled for public release on July 27. If that release happens as described, the Alibaba-backed Chinese AI lab says Kimi K3 will be the largest open-weight model available, putting another low-cost Chinese model into direct comparison with proprietary systems from OpenAI and Anthropic.

The company did not disclose training cost, compute used, revenue impact, customer count or headcount tied to the launch. The main disclosed number is model scale: 2.8 trillion parameters. Moonshot said in a blog post that Kimi K3 remains behind OpenAI’s GPT-5.6 Sol and Anthropic’s Claude Fable 5 in some areas, while its own tests show the model close to those systems on several tasks.

Third-party rankings are doing much of the work in the early narrative. Artificial Analysis has tested Kimi K3 and placed it just behind the top proprietary models on its Intelligence Index and in real-world work evaluations, according to the report. Arena.ai ranked Kimi K3 first on its frontend development leaderboard, ahead of Claude Fable 5, and said the model moved up 17 places from Moonshot’s prior Kimi K2.6 release.

Arena Chief Executive Anastasios Angelopoulos said on X that the release may be one of the year’s most significant and argued that open-source Chinese models had overtaken U.S. models in frontend coding. That is a broad claim from a benchmark operator, not a market fact. It is still a useful signal for founders and engineering leaders watching whether open-weight systems can absorb more developer workflows now dominated by closed APIs.

Software development is the target workload

Moonshot is positioning Kimi K3 for long-running autonomous software engineering tasks. The company said the model is built to examine large codebases, work with programming tools and complete multistep assignments toward a defined goal.

The model also uses visual feedback, according to Moonshot. The company described a system in which Kimi K3 can inspect screenshots, change code and then check the visible result. Moonshot said that approach is suited to game development, user interface work and computer-aided design.

Moonshot released demos showing what it said were projects created by Kimi K3 in a browser, including a 3D open-world game built with Three.js, WebGPU and GPU Compute. The company also showed a Long March 10 rocket launch and return simulation, and a Game Boy Advance emulator. Those demos are company-produced, and the public weights were not yet available at announcement time.

Pricing undercuts closed-model APIs

Kimi K3’s API documentation lists pricing at 30 cents per 1 million input tokens when there is a cache hit and $3 per 1 million input tokens without a cache hit. Output tokens, including reasoning tokens, are priced at $15 per 1 million. Moonshot says those prices apply regardless of context length.

The pricing is below the Western models cited in the report. Claude Fable 5 is listed at $1 per 1 million input tokens and $50 per 1 million output tokens, while GPT-5.6 Sol is listed at 50 cents per 1 million input tokens and $30 per 1 million output tokens. For buyers running high-volume coding agents, that gap is the part of the launch most likely to get procurement attention before broader trust questions are settled.

The comparison invites the obvious DeepSeek precedent. DeepSeek’s R1 release in January 2025 showed that a Chinese lab could put price pressure on U.S. frontier-model vendors, contributing to a sharp selloff in U.S. technology stocks and renewed security scrutiny in Washington.

Constellation Research analyst Holger Mueller said Kimi K3 stands out because of its scale, open weights, multimodal visual feedback and lower price. He also said autonomous software development is a sensible target because AI adoption is already visible there, and model makers are competing for developer workloads.

The release will also feed existing arguments over export controls, model distillation and the security treatment of frontier systems from China. Anthropic previously accused Moonshot, DeepSeek and MiniMax of using Claude to train competing models through distillation, an allegation Moonshot has faced before Kimi K3’s public weight release. The Trump administration has described model distillation as an adversarial approach, though the practical enforcement path remains unclear.

This story draws on original reporting from SiliconANGLE.