Moonshot AI Ships Kimi K3, Calling It the Largest Open-Weight Model Yet

Beijing-based Moonshot AI released Kimi K3 on July 16, a 2.8-trillion-parameter model the company describes as the largest open-weight AI system in the world. The launch lands days before the 2026 World Artificial Intelligence Conference in Shanghai and marks a sharp comeback for a lab whose standing had slipped since DeepSeek's rise. Benchmarks published by the company put K3 within striking distance of the strongest proprietary systems from OpenAI and Anthropic, a claim the open-source community is now racing to verify.

The release is more than a spec sheet. It is a statement that the frontier of open AI has moved to China, and that the gap between open-weight and closed models, long assumed to be permanent, is closing fast enough to unsettle labs on both sides of the Pacific.

The numbers that matter

Kimi K3 carries 2.8 trillion total parameters, roughly 75% larger than the 1.6 trillion Moonshot cites for DeepSeek's V4 Pro. It runs a 1-million-token context window, meaning it can hold the equivalent of several full books in a single prompt, native visual understanding, and an always-on reasoning mode the company brands as "thinking mode." Those are headline figures, but the architectural choices underneath are the more interesting story.

Two techniques developed in-house do the heavy lifting. Kimi Delta Attention is a hybrid linear attention mechanism that the company says enables up to 6.3 times faster decoding in million-token contexts, the kind of speedup that matters precisely because long context is where most models bog down. Attention Residuals, described as a drop-in replacement for standard residual connections, deliver about 25% higher training efficiency for under 2% added compute. Both were published as open research before the model shipped, a pattern that fits the open-weight positioning.

Benchmarks and the open-source claim

Moonshot points to a score of 91.2 out of 100 on BrowseComp, a benchmark for long-horizon, high-difficulty web research tasks, as evidence the model competes at the frontier. Independent analysts have noted that K3 also places at or near the top of open-weight systems on coding and agentic evaluations, though some caution that the scores reflect a degree of benchmark tuning, a common practice across the field.

The open-weight label is the part competitors will scrutinize hardest. Moonshot says full model weights will be released by July 27, along with a technical report covering architecture, training, and evaluation. Until that drop, K3 is available through the company's API and its consumer site, where users can sign up with a Google account or phone number and no credit card. The distinction matters: a model you can chat with is a product; a model whose weights you can download and run is infrastructure, and only the second claim earns the open-source title.

Pricing and developer reach

On the API side, K3 is priced at $3 per million input tokens and $15 per million output tokens, with cached input tokens falling to $0.30 per million. That lines up with mid-tier Western offerings while aiming at top-tier performance, and a promotional rebate through August 12 returns up to 30% in vouchers on API credits of $1,000 or more. The model is compatible with the OpenAI SDK, which lowers the barrier for teams already building on OpenAI or Anthropic toolchains and makes a switch largely a change of endpoint.

For developers, that compatibility is the quiet selling point. A shop that already wraps the OpenAI client can point it at Kimi K3 with a few lines changed, then benchmark the two side by side on their own workloads. The open-weight release, when it lands, takes that a step further: teams can host the model themselves, tune it, and keep data on their own hardware, options that closed APIs do not allow.

Why this lands now

The timing is not accidental. A launch ahead of the Shanghai AI conference maximizes press attention in China and frames the country's open models as co-equal with Western frontier systems. For Moonshot specifically, it is a reset. The lab's market position had eroded over the past eighteen months as DeepSeek captured mindshare with cheaper, capable models. K3 is the answer: bigger, openly licensed, and benchmarked against the best.

The broader signal is about the open-weight tier as a whole. For most of the last two years, the assumption was that the best systems would stay behind paid APIs while open models trailed by a generation. K3, whatever its final independent scores, argues that the trailing edge has narrowed to a gap measured in months rather than years. That changes the calculus for enterprises weighing whether to rent intelligence or own it.

What to watch

Three things will decide how much K3 actually moves the field. First, the July 27 weight release and whether the downloaded model matches the API's behavior, because a gap between the two would undercut the open claim. Second, independent replication of the BrowseComp and coding scores on private tasks, not the public leaderboards. Third, real deployment: whether teams adopt K3 for production work or treat it as a benchmark curiosity.

Moonshot has made the bold claim. The next two weeks, as weights ship and the community puts them through its own tests, will show whether Kimi K3 is the open frontier it claims to be or simply the loudest entry in a crowded race. Either way, the bar for open-weight models just moved, and the labs that assumed a permanent lead should take note.

The geopolitical weight of an open model

A model anyone can download is also a model no export control can easily contain. The US has spent two years tightening curbs on advanced chips and, at times, on model weights themselves, betting that scarcity keeps the frontier on its side of a line. An open-weight release from a Chinese lab undercuts that logic: once weights are public, the constraint shifts from who may build the model to who can run it, and running a 2.8-trillion-parameter system is a hardware problem, not a permission one. That is why K3 lands as a policy event as much as a product launch.

The local deployment reality

Running K3 at home is possible in theory and punishing in practice. A model this size needs serious accelerators and memory, the kind of kit that lives in a datacenter rather than a closet. Early adopters talk about clusters of high-memory machines, not a single box, and inference costs that only make sense for teams with steady volume. The open-weight release democratizes access in principle; in practice it favors those who already own the iron. The interesting deployments will come from operators who pair the weights with efficient serving stacks and treat the model as infrastructure, not a toy.

Our AI section follows the frontier model race as it unfolds, and Semiconductors covers the chips that run these systems.