Open-weight release and competitive performance
Moonshot AI has released the full weights of Kimi K3, making it the largest open-weight model publicly available at 2.8 trillion parameters, built on a Mixture-of-Experts architecture with a 1 million-token context window and native vision understanding. Internal and third-party testing reported by industry coverage places the model close to leading closed systems: Kimi K3 scored 88.3 on Terminal-Bench 2.1, compared with 88.8 for GPT-5.6 Sol and 88.0 for Claude Fable 5, and outperformed both on Moonshot's reported BrowseComp results. The release intensifies competition with OpenAI, Anthropic and Google by letting developers download, fine-tune and deploy the model on private infrastructure rather than access it only through a hosted API.
API pricing undercutting US frontier labs
Kimi K3's API pricing is materially below US competitors, according to published rate cards: $3 per million uncached input tokens and $15 per million output tokens, with cached input at $0.30 per million. Comparable rates cited are $5/$30 for GPT-5.6 Sol, $5/$25 for Claude Opus 5 and $10/$50 for Claude Fable 5. For agentic workloads that consume large token volumes, the headline pricing positions Kimi K3 as the cheapest of the four, though final cost depends on how many tokens each model burns to complete a given task. Coverage framed the pricing as a reason US buyers are already evaluating the model for production use.
White House distillation allegation
The release has moved from a technical milestone to a policy flashpoint. White House science and technology policy chief Michael Kratsios wrote on X that the US "has information that Moonshot AI distilled Anthropic's Fable for the development of its K3 model" and described an internal platform designed to "conduct large scale distillation against U.S. models, allowing them to quickly switch between multiple methods of access to avoid detection." Moonshot rejected the allegation. Anthropic framed the issue as a national-security matter rather than a commercial grievance, arguing that guardrails against bioweapon and cyberattack misuse built into US models may not transfer to copied systems, and cited reports that Claude outputs were used to train rival models at industrial scale via roughly 24,000 fake accounts. Distillation itself is widely used across the industry, including by Nvidia in its Llama Nemotron series, which complicates any blanket prohibition.
Industry split over open-weight restrictions
In late July, more than 20 companies including Nvidia, Microsoft, Meta and Palantir signed a letter warning policymakers against "premature restrictions" on open-weight AI models that would "stifle competition or drive innovation overseas," and describing distillation as a standard technique for "model improvement, evolution, and validation." Box CEO Aaron Levie, a signatory, argued for the opposite risk in walling off models. Notably, OpenAI and Anthropic did not sign, splitting the industry along business-model lines rather than purely national ones. Coverage framed the coalition letter less as a safety argument than as a bet that buyers will continue chasing cheap capability wherever it is hosted.
Chip procurement and export-control questions
Separately, reporting tied Kimi K3's training to Alibaba-supplied Nvidia hardware. Alibaba, one of Moonshot's biggest backers, has pushed portfolio companies onto its cloud, and roughly 20,000 processors in the arrangement are Hopper-generation chips rather than Blackwells, according to people familiar with the setup; several of those sources said the agreement covers H200s, the most powerful Hopper parts, though an Alibaba spokesperson denied supplying H200-powered computing to Moonshot. Kratsios separately accused Moonshot of acquiring Nvidia Blackwell processors illegally, possibly through Thailand, and said the US has not disclosed how many Blackwells it believes Moonshot holds; The Information reported Moonshot wants more Blackwell accelerators for its next model. US rules generally bar Blackwell sales to Chinese buyers but permit cloud-based access, and the Trump administration has issued some H200 licenses to China, though a senior US official said only a "trivial" number of chips have actually shipped under them. Where Alibaba's compute for Moonshot is physically located has not been disclosed.
Congressional inquiry into DoorDash's Chinese-model testing
The policy debate has reached named US users. House Select Committee on China chair John Moolenaar and House Homeland Security Committee chair Andrew Garbarino asked DoorDash CEO Tony Xu to identify every Chinese AI model the company uses, describe any security testing and brief congressional staff, with information requested by August 14 and an in-person briefing by August 21. DoorDash's July engineering post describes DashBench, an internal 105-case code-review evaluation; one configuration used Moonshot AI's Kimi K2.6 as a scout while the company's stated production reviewer uses Claude Sonnet 4.6 and Claude Opus 4.8, which on the 105-case report found 504 real findings at 53.6% weighted recall and 87.0% weighted precision at a stated $3.91 per pull request. DoorDash said it supports American AI leadership and will engage with the committees, but as reported did not yet provide the requested model inventory or describe its data-handling arrangements. The letter places DoorDash within a broader inquiry into US companies' use of China-developed models rather than treating it as the sole target.
Share this article
