Chinese AI startup Moonshot released Kimi K3, a large language model it says approaches the capabilities of frontier US labs such as Anthropic. Built by the Beijing-based, Alibaba-backed company, K3 has 2.8 trillion parameters, which Moonshot calls the largest open-weight model ever released and the first “3T-class” open model. Parameters roughly measure the size of a model’s neural network, with more generally enabling greater capability, though the relationship is not absolute.
Crucially, K3 is open-weight, meaning it can be downloaded and modified freely, with full weights due by July 27. It offers a one-million-token context window and native handling of images and video. The launch was timed just ahead of the World Artificial Intelligence Conference in Shanghai.
The performance claims are striking but require caution. On Moonshot’s own benchmarks, K3 beats Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5 on most coding and agentic tests, while still trailing Anthropic’s more powerful Fable 5 and OpenAI’s GPT-5.6 Sol on several.
These are self-reported figures, and independent verification is only beginning. Early results from the firm Artificial Analysis largely corroborate that K3 sits in the Opus 4.8 tier, placing it second on one agentic knowledge-work benchmark ahead of GPT-5.6 Sol but behind Fable 5, though it noted K3’s hallucination rate rose versus its predecessor. Developer Simon Willison summarized the independent read: K3 mostly beats Opus 4.8 and GPT-5.5 while losing to Fable 5 and Sol.
Notably, K3 also signals the end of ultra-cheap Chinese AI. At about $3 per million input tokens and $15 per million output, it is roughly five times pricier than the K2 family it replaces and comparable to Western mid-range models like Claude Sonnet 5. On a per-task basis it runs around $0.94, similar to GPT-5.6 Sol and about half the cost of Opus 4.8, though far above open rivals like DeepSeek. The parameter count itself is a pointed detail: at 2.8 trillion it exceeds industry estimates that Opus 4.8 holds 1.5 to 2 trillion, figures Anthropic has never disclosed, so the comparison rests on speculation.
China Closes the Gap
K3’s arrival challenges the industry assumption that Chinese models trail US ones by eight to 12 months. It follows Z.ai’s GLM-5.2, which venture capitalist Marc Andreessen called the first Chinese model to match American labs “with no compromises,” and a broader wave in which companies from Silicon Valley to Europe are switching to cheaper Chinese models to cut their bills.
The open-weight strategy is the sharper threat: while US leaders like Anthropic and OpenAI keep their frontier models closed and are raising prices, with Anthropic’s Opus 4.8 set for a 50% increase in September, Chinese labs are giving comparable capability away to run on a company’s own servers. That combination of closing capability and open access puts real pressure on the closed, premium model that funds US labs.
The Distillation Shadow
The launch reopens a bitter dispute. Anthropic accused Chinese labs in February of “industrial-scale distillation attacks,” the practice of training a cheaper model on a more advanced system’s outputs to replicate its performance without the same compute, and later named operators tied to Alibaba’s Qwen.
Chinese companies have pushed back, calling such claims an excuse to protect a US monopoly. The tension is unresolved and colors how K3’s rapid progress is read: US labs frame it partly as the fruit of copying, while Chinese developers frame it as genuine, efficient engineering.
What is not disputed is the commercial gap, with Moonshot raising funds at about $31.5 billion and DeepSeek near $71 billion, still tiny beside Anthropic’s roughly $965 billion valuation. For now, K3 is best understood as a real narrowing of the capability gap paired with unverified “beats everything” claims, a distinction worth holding until independent benchmarks and the open weights arrive.
