Chinese artificial intelligence lab Moonshot AI has announced the imminent release of Kimi K3, a 2.8-trillion-parameter large language model that it says is the largest open-weight system ever built. According to Reuters, the Wall Street Journal, and the South China Morning Post, the model lands just behind the very best proprietary systems from OpenAI and Anthropic on independent benchmarks while undercutting them sharply on price — a combination that observers are already comparing to DeepSeek's market-rattling debut. The weights are scheduled to be released to the public on July 27, making the model freely available for anyone to download and build upon.
The launch arrives amid a wave of breaking AI news underscoring how quickly Chinese labs have closed the gap with their American counterparts. Backed by Alibaba Group Holding, Moonshot said in a blog post that Kimi K3 still trails OpenAI's GPT-5.6 Sol and Anthropic's Claude Fable 5 in some areas, but the company's internal tests show it sitting "extremely close behind" those models on several key tasks. For context, Anthropic only released Fable 5 last month, and OpenAI's GPT-5.6 debuted just a week earlier.
What the benchmarks show
Independent evaluation firm Artificial Analysis has already put Kimi K3 through its paces. According to SiliconANGLE's coverage, the model places just behind the top proprietary models on Artificial Analysis's Intelligence Index and in real-world work evaluations. But on Arena.ai's front-end development leaderboard, Kimi K3 actually ranks above GPT-5.6 Sol and Claude Fable 5 — a striking result for an open-weight system. That ranking puts it 17 places above Moonshot's previous release, Kimi K2.6.
Arena Chief Executive Anastasios Angelopoulos made a particularly bold claim in a post on X, calling the release what may be "the single biggest release of the year," and suggesting it could represent the moment China finally surpassed the United States in model capability. Axios and Fortune echoed that framing, reporting that Kimi K3 "stuns the AI world" and pushes Chinese AI into frontier-level territory.
Built for autonomous software engineering
In its blog post, Moonshot explained that Kimi K3 was designed primarily for long-running autonomous software development tasks. The model can analyze large codebases, coordinate external programming tools, and execute multistep tasks to achieve an end goal — the kind of agentic workflow that has become a major battleground for AI labs in 2026.
A standout feature is what Moonshot calls a "vision-in-the-loop" system. Kimi K3 can examine screen captures, modify code, and then inspect the resulting visible output to verify its own work. Moonshot says this makes the model especially well-suited for games development, user-interface design, and computer-aided design.
To demonstrate those capabilities, Moonshot posted a video showing a 3D open-world game that Kimi K3 reportedly built entirely inside a web browser using Three.js, WebGPU, and GPU Compute. The model procedurally generated the environment and used external tools to create a 3D rider and horse. Additional demos included a simulation of China's Long March 10 rocket launch and return, and a Game Boy Advance emulator.
Dramatically cheaper than Western alternatives
The pricing is where Kimi K3 could prove genuinely disruptive. According to the model's API documentation, Kimi K3 costs 30 cents per 1 million input tokens with a cache hit, and $3 without. One million output tokens, including reasoning, cost $15. Those rates apply regardless of context length.
That makes Kimi K3 substantially cheaper than its Western competitors. Anthropic's Claude Fable 5 charges $1 per 1 million input tokens and $50 per 1 million output tokens, while OpenAI's GPT-5.6 Sol costs 50 cents per 1 million input tokens and $30 per 1 million output tokens, according to SiliconANGLE's comparison.
Another DeepSeek moment?
The release has prompted immediate comparisons to DeepSeek, whose R1 model triggered roughly $1 trillion in losses across major technology stocks when its capabilities became clear in January 2025. Former White House AI policy adviser Sriram Krishnan summed up the mood on X, calling Kimi K3's debut "a big moment, with multiple implications for the entire industry."
The launch also lands against a backdrop of escalating scrutiny. Earlier this year, Anthropic accused Moonshot, DeepSeek, and MiniMax of violating its rules by extracting the capabilities of Claude to train their own models — a process known as distillation. The Trump administration has since labeled distillation an "adversarial" approach and promised to crack down on it, meaning Kimi K3's open-weight release could reignite debate over export controls, model security, and whether restrictions on Chinese AI firms are having any measurable effect.
Neowin reported the broader takeaway plainly: with Kimi K3, open AI models have "finally caught up" with their proprietary US-based rivals. Whether the market reacts as violently as it did to DeepSeek remains to be seen, but the underlying claim — that open-weight Chinese models now compete at the frontier — is becoming harder to dispute.
For Moonshot, the bet is that freely releasing a frontier-grade model's weights will accelerate adoption and cement China's position in the global open-source ecosystem. The world will get its first full look when the weights go live on July 27.
Stay ahead of the AI curve
Want more coverage like this? Bookmark AI Buzz Wire for the latest AI developments, model releases, and analysis from across the industry.
Read more AI news →
