Kimi K3 Beat Claude Fable 5 at Front-End Coding. The Weights Aren't Out Yet.

Simon Willison has a ritual. Every time a major lab ships a new model, the software developer and AI blogger behind the widely used llm command-line tool asks it to draw an SVG of a pelican riding a bicycle. It is a small, deliberately silly test, useless for grading real intelligence, but good at revealing how a model behaves on a task nobody trained it to memorize.

On July 16, he pointed the test at Moonshot AI's newest release. The model thought for 13,241 reasoning tokens before drawing a single pixel.

The bill for one pelican: 25 cents.

That is an unusual amount of money to pay a Chinese AI lab for a cartoon bird. It is also, in miniature, the story of Kimi K3.

Moonshot Shipped the Largest Open-Weight Model Anyone Has Released

Kimi K3 is a 2.8 trillion parameter model built by Moonshot AI, a Beijing lab founded in 2023 that has spent the past year building a following among Western developers. Moonshot announced the model on its official blog on July 16, describing it as the first "open 3T-class model," a rounding-up from 2.8 trillion that Willison flagged in his own write-up.

BofA Securities analyst Alex Liu did not round up. In a note to clients, he described K3 plainly as "the largest open-weight model released globally to date."

The architecture is what makes that possible. Kimi K3 is a sparse mixture-of-experts model, meaning it does not run all 2.8 trillion parameters for every request. It activates just 16 of 896 experts per token, roughly 1.8% of the full network. That design let Moonshot build something 75% larger than the previous open-weight record holder, DeepSeek's 1.6 trillion parameter V4 Pro, without a matching jump in the compute needed to run it.

Two other choices did the rest of the work. Kimi Delta Attention, a hybrid linear-attention mechanism, lets the model decode up to 6.3 times faster as conversations stretch toward its 1 million token context window. A second technique, called Attention Residuals, reportedly improves training efficiency by about 25%, for under 2% in added compute cost. Neither claim has been independently verified yet.

Yutong Zhang, Moonshot's president, described the constraint behind those choices at the World Economic Forum earlier this year: "We knew we didn't have the luxury to simply scale up compute. That forced us to focus on fundamental research and efficiency." US export controls have limited Chinese labs' access to the most advanced chips, pushing companies like Moonshot toward efficiency gains instead of brute-force scaling.

Kimi K3 also processes images and video natively, and unlike most models on the market, it currently ships with only one reasoning setting: maximum. There is no faster, cheaper mode to switch to. Willison's own testing suggested an 85-token hidden system prompt sitting behind the API. Kimi K3 would not tell him what was in it.

The Benchmarks Put It Ahead of Claude on Its Own Turf

Moonshot's own numbers, later checked in part by the independent evaluator Artificial Analysis, put K3 behind only Claude Fable 5 and GPT-5.6 Sol on overall capability. Those two systems represent the current ceiling of what American labs have shipped, so trailing them and nothing else is a real result, not a rounding error.

The clearest evidence came from Arena, the evaluator that runs blind head-to-head tests between models with the model identities hidden from the developers judging them. Developers preferred Kimi K3 over every other tested system, including Fable 5 and GPT-5.6 Sol, for front-end coding. On Arena's broader text leaderboard, K3 also outranked Claude Opus 4.8, the model that held Anthropic's frontier position until Fable 5 replaced it in June.

The pattern holds across individual benchmarks, though it is not a clean sweep:

On GDPval-AA v2, a benchmark built around real occupational tasks across dozens of industries, K3 finishes third: a genuine result, and a sharp jump from Kimi K2.6's score of 1,190 on the same test. On GPQA Diamond, a graduate-level science exam, K3's 93.5% is the strongest score of any open-weight model at launch, ahead of Chinese rival GLM-5.2's 91.2%, though GPT-5.6 Sol still edges it out overall.

The cost line is where the comparison sharpens. Artificial Analysis measured K3 landing close to GPT-5.6 Sol on price per task and at roughly half of what Claude Opus 4.8 costs to run the same workload. It is still pricier than open-weight peers like GLM-5.2 and DeepSeek V4 Pro, but for a model within range of the American frontier, that is a meaningfully different tier than Chinese labs have offered before.

Not every number moved in Kimi's favor. Its accuracy on Artificial Analysis's Omniscience benchmark rose from 33% to 46% compared to K2.6, but its hallucination rate climbed too, from 39% to 51%. K3 gets more answers right and fabricates more of them along the way.

Wall Street Is Already Repricing the Race

The reaction on trading desks arrived almost as fast as the model itself. AI and semiconductor stocks sold off in the hours after the announcement.

"K3 has received positive feedback globally, signaling an all-round catch-up of Chinese LLMs with US leaders in model size, performance, and pricing," Morgan Stanley analyst Gary Yu wrote in a note to clients. "We do not view K3 as an overnight miracle but rather as the result of cumulative progress across China's AI model industry."

Bernstein analyst Robin Zhu called K3 "a home run" and suggested its arrival could sharpen the political fight already underway in Washington over Chinese AI.

The reaction was not confined to trading desks. Mozilla CTO Raffi Krikorian framed the stakes bluntly to Axios: "Right now, it's a U.S. versus China question." He argued that American AI executives lobbying Washington against open-weight models would have little reason to do so unless they saw Chinese open models as a genuine competitive threat.

The Price Broke Moonshot's Own Playbook

For two years, Chinese AI labs have competed largely on being dramatically cheaper than their American counterparts. Kimi K3 abandons that approach.

Token for token, K3 now costs exactly what Anthropic charges for Claude Sonnet 5. BofA's Liu called it the most expensive model any Chinese lab has released, a sharp break from Z.ai's GLM-5.2 and DeepSeek's V4, both still priced for volume.

Then there is the branding problem. Moonshot calls K3 open, and says it will publish the full model weights on July 27. Until then, the only way to run it is through Moonshot's own hosted API and apps, the same closed-door arrangement Chinese labs have spent two years positioning themselves against.

Moonshot is also reportedly raising a new funding round that would value the company at approximately $31.5 billion, according to TechCrunch.

The company was valued at just $20 billion two months earlier, when it closed its previous round. Two other outlets have since cited similar figures for the new valuation, though the round has not closed and Moonshot has not confirmed the number publicly.

Not Everyone Is Convinced This Changes the Race

Shakeel Hashim, editor of the AI policy newsletter Transformer, argued the panicked reaction misreads what actually happened. By Moonshot's own admission, he wrote, K3 "is not at the frontier," meaning it likely lacks the dangerous cyber capabilities that led Washington to restrict access to Anthropic and OpenAI's most powerful systems in the first place. Releasing a model that is not at the frontier costs China little and buys real goodwill.

There is a harder problem sitting underneath his argument. As Axios noted within hours of launch, Kimi has been reachable only through Moonshot's own hosted API and apps so far, and the company does not plan to release the weights until July 27. That means every benchmark discussed above, including the ones where K3 loses to Fable 5, comes from Moonshot's own testing or from third parties working through Moonshot's hosted access, not from researchers running the model independently on their own hardware.

He also expects the openness to be temporary. "That does not mean China will keep open-sourcing models forever," he wrote, pointing to Reuters reporting that Beijing is already weighing restrictions on overseas access to its most capable models.

Real-world reaction inside China has been messier than the benchmark charts suggest. 36Kr, the Chinese business outlet, reported developer feedback describing K3 as running two to three times slower than rival models, burning through subscription quotas quickly, and occasionally failing on coding tasks that Claude Opus handled cleanly. The harshest assessment in that coverage: the launch "reeks of over-the-top marketing," with real-world performance lagging the claimed top-three global ranking.

Then there is the distillation question hanging over every Chinese release this year. Anthropic has formally accused Moonshot, alongside DeepSeek, Alibaba and MiniMax, of training on millions of exchanges with Claude to build competing systems. Moonshot has not addressed that claim specifically in relation to K3.

The bigger pattern is one Let's Data Science covered yesterday: Chinese open-weight models captured 41% of Hugging Face downloads this spring, more than American models for the first time, and now fill the top spots on OpenRouter's popularity leaderboard. Kimi K3 does not create that trend so much as raise its ceiling, showing that the labs driving those download numbers can now compete on raw capability, not just price.

The Bottom Line

Strip away the framing on both sides and the facts are not actually in dispute. Moonshot shipped a 2.8 trillion parameter model that beat two of the best closed models in America in a blind front-end coding test, priced it like a premium American model instead of a discount one, and will not let anyone download it until July 27.

Whether that adds up to China erasing America's AI lead, as several headlines claimed within hours of release, or a strong but unverified result from a lab with every incentive to look as close to the frontier as possible, depends on what happens after July 27. Independent researchers running K3 on their own hardware, rather than through Moonshot's API, will settle far more of this argument than any benchmark table can.

Willison's pelican cost 25 cents to draw. Figuring out whether Kimi K3 is actually worth its new price tag will cost the rest of the industry a great deal more than that.

Sources

• Kimi K3, and what we can still learn from the pelican benchmark — Simon Willison's Weblog, July 16, 2026
• Moonshot AI Releases Kimi K3: A 2.8 Trillion Parameter Open MoE Model With Kimi Delta Attention and 1M Context — MarkTechPost, July 16, 2026
• Moonshot's Kimi K3 pushes Chinese AI into Fable-level territory — Fortune, July 16, 2026
• China's open-weight Kimi model stuns AI world with frontier-level results — Axios, July 16, 2026
• China just erased America's AI lead — Axios, July 17, 2026
• Kimi's open model K3 nears GPT-5.6 Sol and Fable 5 while signaling the end of super cheap Chinese AI — The Decoder, July 16, 2026
• Kimi K3 Beats Fable 5, GPT 5.6 On Some Benchmarks In Frontier-Level Performance — OfficeChai, July 16, 2026
• Claude Fable 5 vs Kimi K3: Benchmarks, Pricing, Speed — BenchLM.ai, data verified July 16, 2026
• Moonshot Unveils 2.8 Trillion Parameter 'Kimi K3', Mounting a Direct Challenge to Anthropic — BigGo Finance, July 17, 2026
• Kimi K3 AI breakthrough: What Wall Street analysts say about China's OpenAI threat — Yahoo Finance / Investing.com, July 17, 2026
• Kimi K3 is no reason for China panic — Transformer News, July 17, 2026
• The High-Priced Kimi K3, the Cash-Strapped Moonshot AI — 36Kr, July 17, 2026
• Moonshot's upcoming Kimi 3 is expected to close the gap with Anthropic's Opus 4.8 — TechCrunch, July 16, 2026