Moonshot AI just put out the biggest Chinese open-source model ever released, and it topped Claude Fable 5 at writing scripts.

Towards AI’s Writing Elo—a benchmark where models write real scripts judged blind against published versions, scored using the same Elo system that ranks chess players—put Kimi K3 at 2,840, above Fable 5 (max) at 2,760. That’s a ranking Anthropic’s team has historically dominated.

K3 also claimed the top spot on Arena AI’s Frontend Code Leaderboard—a ranking built from thousands of pairwise human votes on code generation tasks, again Elo-scored—with 1,679 against Fable 5’s 1,631. First place in six out of seven frontend domains.

The Artificial Analysis Intelligence Index—a score built from nine independent evaluations covering coding, reasoning, agentic work, and knowledge, rated 0 to 100—puts K3 at 57, with Claude Fable 5 at 60, GPT-5.6 Sol at 59, and Claude Opus 4.8 at 56. That places K3 as the third-most capable model on the composite, with Fable 5 beating it just by 3%.

If you want to have an idea of what it can do, this is a zero-shot result of a prompt asking the model to build an iOS clone. For comparison, this is the best approximation shared on social media using GPT 5.6 Sol and a much elaborate prompt.

What this thing actually is

K3 packs 2.8 trillion parameters—the numerical values that store a model’s knowledge—in a mixture-of-experts architecture. Mixture of experts splits those parameters into 896 “expert” subnetworks and activates only a fraction for any given task. That’s how you get frontier-level intelligence without melting the server room.

“It is the world’s first open-source model in the 3-trillion-parameter class, designed for frontier intelligence scenarios including long-horizon coding, knowledge work, and reasoning,” Moonshot AI says. That’s not marketing theater: DeepSeek’s V4-Pro tops out at 1.6 trillion parameters, Moonshot’s own K2 at one trillion. K3 roughly doubles the nearest open-weight competitor on the size chart.

It comes with a one-million-token context window—tokens are the basic unit of information an AI processes, about three-quarters of a word each—native image and video understanding, and always-on reasoning.

Two architectural techniques underpin the efficiency gains. Kimi Delta Attention speeds up decoding for long sequences—up to 6.3x faster at million-token contexts. Attention Residuals routes information selectively across model layers rather than accumulating it uniformly, adding about 25% training efficiency at under 2% extra compute cost—together yielding roughly 2.5x better scaling efficiency than K2.

Benchmarks are nice, Prices are nicer

Kimi K3 costs $3 per million input tokens and $15 per million output tokens—the same rate as Claude Sonnet 5, Anthropic’s mid-tier model. The difference is that Sonnet 5 is Anthropic’s middle-ground offering; K3 is sitting three points below Fable 5 on the Artificial Analysis composite. Per task across that nine-benchmark suite, K3 runs $0.94 versus $1.04 for GPT-5.6 Sol and $1.80 for Opus 4.8.

In other words, this model offers top of the line performance at mid-tier level prices.

As Decrypt covered in May, the pricing gap between Chinese and American frontier AI ran 15–30x earlier this year. K3 doesn’t undercut at DeepSeek rates—it prices like a Western mid-range model—but delivers near-frontier performance at that level. For teams building on the API this represents a major cost improvement.



If Anthropic goes ahead with its intentions of making Fable 5 available only via API, K3 becomes the nearest open-weight alternative to whatever model currently sits second in the industry—at half the per-task cost of Opus 4.8. That’s the scenario benchmark chasers are already running the math on.

K3’s launch is the argument U.S. chip export controls advocates don’t want to have. The U.S. restricted Nvidia’s H800 GPUs from export to China in late 2023; Moonshot confirmed it trained earlier models on those chips. K3’s own benchmark documentation references H200s and what the company calls “a GPGPU from an alternative vendor”—widely interpreted as Huawei Ascend hardware—without specifying where that hardware sits.

Moonshot AI president Yutong Zhang framed the constraint directly at Davos this year, per Silicon Republic: “We knew we didn’t have the luxury to simply scale up compute… That forced us to focus on fundamental research and efficiency.” Bank of America analysts, in a note after the launch, wrote that K3 proves “pre-training scaling, paired with architectural innovation, can still deliver step-change gains for flagship Chinese models” under those constraints.

Moonshot is one of the so-called AI Tiger startups that have collectively shifted the global model landscape without access to the chips Washington said they’d need. Whether that’s an argument for tighter export controls or an argument that they don’t work is a policy question Washington hasn’t settled.

The asterisk you should read

K3’s hallucination rate on AA-Omniscience—a benchmark that measures how often a model confidently fabricates an answer it doesn’t know—jumped from 39% to 51% compared to predecessor K2.6. More correct answers overall; more made-up ones too. The model also acknowledges in its own documentation that it can be “excessively proactive,” making unexpected decisions on a user’s behalf during long autonomous tasks.

For teams that ran the Kimi K2.6-based tooling and want to upgrade, K3 is a meaningful step up on most fronts—but that hallucination delta is worth stress-testing before you trust it with anything that needs to be accurate.

If you want to try it for free, you can. It’s available on Kimi’s official website. But good luck: The servers are so packed that tasks get interrupted constantly due to traffic constraints, making it barely usable. A better alternative is to either pay for a subscription or use it over an API.

Weights will be released on July 27. Those will be available for big enterprises and businesses. No domestic GPU, no matter how big, is currently able to handle a model this size.