Kimi K3 Is China’s Biggest Swing Yet at the AI Frontier, And Wall Street Felt It Immediately

Moonshot AI dropped a 2.8-trillion-parameter model on the world on July 16, called it open, and watched semiconductor stocks bleed for the rest of the week. Kimi K3 is the largest open-weight model anyone has ever shipped, and depending on which leaderboard you trust, it is either the fourth-best model on Earth or the outright best at building web interfaces. Both readings are defensible. Neither is comfortable if you run a closed-weights lab in San Francisco.

What Moonshot Actually Shipped

K3 is a mixture-of-experts system with 2.8 trillion total parameters, of which 16 out of 896 experts fire on any given token. That’s a much sparser, wider routing shape than K2.6, which ran 384 experts with eight routed and one shared. The context window is the headline spec: 1,048,576 tokens, four times what the previous generation offered, and priced flat across the whole window rather than stepped by tier.

Vision is native rather than bolted on — text, image, and video go in, text comes out. Reasoning is always on, running at max effort by default with no toggle to dial it down at launch. Moonshot calls this “thinking mode,” and it is both the model’s biggest strength and the source of the loudest early complaints.

Two in-house architectural pieces do the heavy lifting. Kimi Delta Attention is a hybrid linear attention scheme built to keep very long sequences tractable. Attention Residuals is a drop-in replacement for standard residual connections that lets later blocks pull from earlier representations rather than inheriting a strictly sequential stream. Both were published as open research before the model landed, which is a nice change from the usual pattern of shipping first and explaining never. The whole thing runs on MXFP4 weights with MXFP8 activations, and Moonshot’s own deployment guidance recommends a supernode with at least 64 accelerators. Nobody is running this on a workstation.

The Spec Sheet

The Benchmarks Are Real, and They’re Messier Than the Headlines

Start with the win Moonshot most wants you to see. K3 opened at number one on Arena’s frontend coding leaderboard with a score of 1,679, ahead of Claude Fable 5 at 1,631 and GPT-5.6 Sol at 1,618. That is human preference voting on real UI generation tasks, not a synthetic score, and it is the first time an open model from Beijing has taken the board the American labs had owned for a year.

Zoom out and the picture flattens. Artificial Analysis scored K3 at 57 on its Intelligence Index, ranking it fourth among 189 models — behind Fable 5 and two GPT-5.6 Sol reasoning configurations, ahead of Opus 4.8, GPT-5.5 at extended effort, and Sonnet 5. Vals AI slotted it second of 38 on its composite index at 74.7%. So: top five, not top one.

The coding numbers, in Moonshot’s own reporting, look like this:

• DeepSWE: 67.5 with the KimiCode harness, 67.3 with the standard mini-SWE-agent setup
• Program Bench: 77.8 raw pass rate
• Terminal-Bench 2.1: 88.3
• SWE Marathon: 42.0
• FrontierSWE: 81.2, against Fable 5’s 86.6
• BrowseComp: 91.2%, a record for web-agent work
• GPQA-Diamond: 93.5%

Read those margins carefully. K3 trails GPT-5.6 Sol by half a point on Terminal Bench and leads it by two-tenths on Program Bench. Gaps that small are noise dressed up as results — harness choice, retry policy and tool configuration matter more than the model at that resolution. Moonshot’s own table mixes KimiCode, Claude Code, Codex and mini-SWE-agent results, which makes clean apples-to-apples comparison basically impossible until independent evaluators redo the work on a single harness.

To Moonshot’s credit, the company said out loud in its launch post that K3 still trails the top proprietary systems overall. That candor is rarer than it should be, and it makes the rest of the claims easier to take seriously.

The Price Is the Story Nobody Expected

Here is where K3 breaks the script. Chinese labs have spent two years winning on cost. K3 doesn’t. At $3 per million input tokens and $15 per million output, it is the most expensive model any Chinese lab has ever put on an API — roughly triple K2.7 Code’s input rate and close to four times its output rate.

It’s still cheaper than Fable 5, and the $0.30 cached-input rate is aggressive for anyone with stable prompt prefixes. Moonshot claims cache hit rates above 90% on typical coding workloads. But the always-on max reasoning eats that advantage fast. Early testers reported burning 13,000+ reasoning tokens on a trivial SVG generation prompt, roughly a quarter per query. That’s a bad ratio for anything high-volume.

The practical read for teams: K3 earns its keep on hard, long, multimodal work — big repositories with screenshots attached, multi-day agent runs, million-token research threads. For shipping day-to-day pull requests, K2.7 Code remains the better per-dollar tool and isn’t going anywhere.

The Open-Weight Asterisk

Every “largest open model ever” headline carries a footnote: as of launch, you could not download it. Moonshot set the full-weight release date for July 27, alongside a technical report detailing the real architecture and training details. Until those files hit Hugging Face, K3 is an API product with an open-weight promise attached.

That date is the one that actually matters. A near-frontier model that enterprises can self-host, inspect, fine-tune, and keep inside their own data residency boundaries doesn’t just close a capability gap — it attacks the subscription business through which American labs convert research leads into recurring revenue. Announcing a big model is easy. Handing it to the people who would otherwise pay per token is the move with teeth.

Markets Didn’t Wait for the Weights

A chip selloff was already underway when K3 landed, and the release poured accelerant on it. TSMC fell 7% on Friday despite posting a 77% jump in quarterly operating profit — a gap between fundamentals and sentiment that tells you exactly what investors were reacting to. SoftBank dropped 9%. Z.ai, a domestic rival, cratered nearly 30% in Hong Kong. The Nasdaq 100 slid 1%, Nvidia lost 1.2% and briefly surrendered its spot as the world’s most valuable company to Apple, and Meta shed over 2.4%.

The timing wasn’t accidental. K3 arrived days before the World Artificial Intelligence Conference in Shanghai, where Moonshot’s booth became the main attraction. For a Beijing company founded by former Google researcher Yang Zhilin — one whose market position had eroded badly after DeepSeek’s rise — it’s a comeback with a lot of leverage behind it. The Kimi chatbot passed $200 million in annualized recurring revenue in April.

It also lands in the middle of an increasingly awkward policy conversation. Three years of tightening export controls on GPUs and lithography gear have not stopped Chinese labs from reaching the frontier, and a House Foreign Affairs Committee hearing on export control gaps two days before launch now reads as unfortunate timing.

What This Means

Kimi K3 did not dethrone anyone. Fable 5 still leads on verified coding and deep reasoning, GPT-5.6 Sol holds its ground, and Moonshot admits both. What K3 did was compress the argument. A model built under export restrictions now sits inside half a point of American flagships on multiple agentic benchmarks, wins outright on frontend code generation and web-agent tasks, and will reportedly be downloadable within days.

If you’re evaluating it, be disciplined. Wire it behind a feature flag on one high-value workflow, log tokens, dollars, and success rate against whatever you’re running now, and don’t assume the launch benchmarks will transfer to your codebase. Treat July 27 as a separate milestone to verify rather than a done deal.

The lead the U.S. holds is still real. It is also measured in months now, not years — and the countdown clock just moved to the open side of the board.