China Just Shipped the Largest Open Model in the World. And It Beat Claude at Frontend Code.

For two years the argument was that open models trail the frontier by about a year, and that the gap is structural. On July 16 a Chinese lab released a 2.8 trillion parameter model, said the weights ship eleven days later, and put it at the top of a blind coding leaderboard. The gap is now measured in weeks.

Here is the direct answer. Moonshot AI's Kimi K3 is real, it will be the largest open-weight model anyone has shipped, and the headline benchmark is accurate but narrow. On Arena.ai's Code Arena WebDev leaderboard, where developers vote blind on which of two anonymous outputs is better, K3 finished first with an Elo of 1679, above Claude Fable 5. On general intelligence it finishes third, behind Claude Fable 5 and GPT-5.6 Sol. Both of those things are true at once, and the interesting part is not who won. It is that a model you will be able to download is now close enough to argue about.

What Actually Happened

Moonshot released K3 on July 16, 2026 as a hosted model, on the API and in the Kimi apps. Ten days later, at the time of writing, the weights are still not public. Independent model trackers still classify K3 as proprietary, and there is no download on Hugging Face. Moonshot's tech blog and a WeChat announcement both name July 27 as the date the full weights and a technical report go up.

So the honest framing for today is this: K3 is a hosted frontier-adjacent model that is about to become the largest open-weight release in history. If you are reading this on the 27th or later, check whether the files actually landed before you plan anything around them. Announced release dates and shipped weights are not the same object.

The Benchmark Everyone Is Quoting, and the Four They Are Not

The claim doing the rounds is that K3 beat Claude Fable 5. It did, on one leaderboard, and it is a good leaderboard. Arena.ai's Code Arena shows developers two anonymous model outputs for the same task and asks them to pick the better one. Identities are revealed only after the vote is cast, which removes brand preference from the result. On the WebDev board K3 took first in six of the seven frontend domains, placing second only in gaming, behind Fable 5. Arena also notes this was a seventeen-place jump from Kimi K2.6, which sat at eighteenth.

That is a genuinely strong result and it is worth taking seriously precisely because it is blind and human-judged. It is also one benchmark, on one category of work. Here is the fuller picture.

Two things stand out. The first is that Moonshot's own launch materials put K3 behind both Fable 5 and GPT-5.6 Sol on overall performance. A lab that undersells its own model relative to an independent blind test is doing something unusual, and it makes the rest of its numbers more credible, not less. The second is that the pattern is consistent: K3 leads on frontend and automated-workflow work, and trails on general reasoning and high-stakes professional tasks. That is a shape, not a fluke. It is also, on the Intelligence Index, comfortably the strongest open-weight model in existence, which is the claim that actually matters.

How a 2.8 Trillion Parameter Model Stays Cheap

The parameter count is the headline and it is also the most misleading number in the release. K3 is a sparse mixture of experts: it holds 896 experts and activates 16 of them per token, roughly 50 billion active parameters. You are not paying to run 2.8 trillion parameters on any given request. You are paying to run something Llama-sized, with a very large library of specialists behind it.

Two architectural changes do the rest of the work. Kimi Delta Attention is a linear attention scheme with a fixed-size memory rather than full token-to-token comparison, used in three of every four attention layers. Attention Residuals let a layer selectively draw on the outputs of earlier layers through block-level attention. Together Moonshot reports roughly a 2.5x improvement in training efficiency over its predecessor, measured as model capability per unit of compute, with the block-level variant matching performance on 20% less training compute. The full technical report is still to come, but the direction is consistent with the pricing.

Worth noting alongside the parameter count: K3 takes text, images, and video as input, not just text. It is a multimodal model, which is not how most of the coverage has framed it.

The last piece is quantization. K3 uses quantization-aware training from the supervised fine-tuning stage onward, with MXFP4 weights and MXFP8 activations, chosen for native support on NVIDIA Blackwell and AMD MI400 accelerators. Training in the format you will serve in, rather than compressing afterwards, is why a 4-bit model can hold quality. It also cuts memory bandwidth requirements by roughly 4x against FP16.

That combination, sparse activation plus trained-in 4-bit plus a cheaper attention scheme, is the entire story of how a model this large is being sold at $3 per million input tokens. It is an engineering answer to a compute constraint, which is worth noting given the constraint Chinese labs are working under.

The Pricing, Honestly Compared

K3 is the most expensive model any Chinese lab has released, which is itself a signal about confidence. Against the frontier it still undercuts significantly, but by less than the headlines suggest.

Read that table carefully. K3 is priced exactly at Claude Sonnet 5's standard rate, not below it, and Sonnet 5 is currently running an introductory $2 and $10 through the end of August, which makes it cheaper than K3 today. The dramatic gap is against Fable 5, which is a different tier of model. "A fraction of frontier pricing" is fair against the top of the range and misleading against the middle of it. It is also a steep rise from Moonshot's own last model: K2.6 was $0.95 and $4.00, so K3 is roughly a 3x price increase generation over generation.

Then there is verbosity, which does not appear on any price list. K3 is genuinely chatty: Artificial Analysis measured 130 million output tokens to run its Intelligence Index against a median of 63 million for comparable reasoning models. Output tokens are the expensive half, so that should erase the discount.

It does not, quite. On the same evaluation K3 burned about 24,000 output tokens per task against GPT-5.6 Sol's 15,000, and still came in cheaper at $0.95 per task versus $1.04. The lower per-token rate absorbs the extra tokens with a little room to spare. So the honest version is not "the savings are an illusion", it is "the savings are real but roughly a tenth of what the sticker implies". That is still the kind of thing you want to measure on your own workload rather than take from a leaderboard, which is the same discipline as judging a coding agent by finished tasks rather than by token price.

"Open Weights" Does Not Mean You Can Run It

This is where most of the excitement quietly deflates. The full K3 weights are approximately 1.4 TB, even with MXFP4 quantization. The practical minimum for self-hosting is an eight-node cluster with eight 80GB GPUs each, about 5.12 TB of VRAM total, leaving headroom for the KV cache and activations. Serving it also needs a mixture-of-experts-aware framework with support for 896-expert routing.

So open weights here means something specific and worth being precise about. It means auditable, forkable, and not dependent on one company's uptime or terms of service. It does not mean runnable on your workstation, or on a single node, or cheaply. For all but a handful of organisations, the practical path to K3 is a hosted endpoint, exactly like a closed model, with the difference being that the weights exist somewhere you could in principle go and get them.

That difference still matters. It is the difference between a supplier and a dependency. But it is not the difference between paying and not paying.

Where It Is Available Today

You do not have to wait for the weights to try it. K3 is already served in three places, and testing it on your own workload is worth more than any leaderboard in this post.

• 01OpenRouter. Search for moonshotai/kimi-k3. Routes to Moonshot's INT4 endpoint, and the fastest way to A/B it against whatever you use now without touching your billing setup.
• 02Moonshot's platform. Model ID kimi-k3, available through the API, the web app, and mobile.
• 03Kimi Code API. Model ID k3, aimed at coding-agent workloads specifically.

What I Would Actually Use It For

The useful question is not whether K3 beats Claude. It is which jobs it should get. Three cases make sense to me on the evidence available.

The second case is high-volume work with a large shared prefix. Cached input at $0.30 per million tokens is the most interesting number in the whole release, more so than the parameter count. If you are running the same long system prompt or the same document corpus across thousands of calls, that price changes the arithmetic of the entire product, not just the line item.

The third is long-context work that genuinely needs a million tokens. Plenty of people claim they need it. Few actually do. But whole-repository analysis, long agentic traces, and large document sets are real, and having a second credible option at that window size is worth something.

What I would not do is migrate wholesale. Model choice is a routing decision now, not an allegiance, and the teams that treat it that way are the ones that keep their options open. If you have not built the habit of moving context between models cleanly, that is the actual prerequisite here, not picking a favourite.

The Honest Part

Four things about K3 that the launch coverage is skipping.

It is not the best model. It is the best open model, and it is first at one category of coding. Moonshot says so itself. The framing of "China caught up" is more satisfying than "China shipped the strongest open model ever made, third overall, at a good price", but the second sentence is the accurate one.

The price advantage is real but small. Not the illusion I expected when I saw the verbosity numbers, and not the 3x the sticker implies either. Roughly 9% cheaper per task than GPT-5.6 Sol on the one end-to-end measurement available. Combined with the single reasoning tier, that is a much narrower economic case than "a fraction of frontier pricing" suggests.

There are operational quirks. K3 requires complete historical reasoning traces in multi-turn sessions, which is a real integration constraint if your harness trims history. It can be overly proactive in resolving ambiguity, which means it guesses instead of asking. And its token accounting looks inflated, with independent testers billed around 95 tokens for a prompt that should cost roughly ten, suggesting a sizeable hidden system prompt on every call.

The strategic point is the one that holds. A lab working under compute export controls shipped a 2.8 trillion parameter model that beats a US frontier model at a real task, and is giving away the weights. Whatever you conclude about the benchmark, that is the part that does not have an asterisk.

The gap that mattered was never the benchmark gap. It was the assumption that you would always have to rent your best model from someone else.

If You Only Do Three Things, Do These

• 01Run K3 on your hardest real task through OpenRouter this week. Ten minutes of your own evidence beats every leaderboard in this article.
• 02Measure cost per completed task, not cost per token. K3 is verbose and has one effort level, so the real saving is far narrower than $3 against $10 suggests.
• 03Read the licence the moment the weights land, before you build on them. Modified MIT is expected, not confirmed.

The thing worth watching is not the leaderboard position, which will change again within a month. It is the release cadence. Kimi K2 to K3 with a 2.5x efficiency improvement, shipped open, priced at Sonnet levels, is a lab that has found something repeatable. The next one arrives sooner than you expect, and the sensible response is to build systems that can swap models rather than systems that assume one.

Questions People Ask

Written July 26, 2026, ten days after the K3 launch and one day before the announced open-weights date. The Arena.ai result is taken from Arena.ai's own published leaderboard announcement. Benchmark scores, architecture details, and the per-task cost figures come from Artificial Analysis and from DeepLearning.AI's technical breakdown in The Batch; the single-effort-level and token-accounting observations are from Simon Willison's independent testing on launch day. Model specifications and the release timeline are from Moonshot AI's launch announcement as reported by Tom's Hardware, VentureBeat, CNBC, and Fortune, with the MXFP4 quantization and self-hosting figures from a community technical overview on Hugging Face. Claude pricing is from Anthropic's own model documentation, current as of this date. Three things are explicitly unconfirmed and flagged as such above: the licence, which Moonshot has not disclosed, the weights themselves, which had not appeared on Hugging Face at the time of writing, and the hallucination rate, which has not been published. The 2.5x training-efficiency claim is Moonshot's own and awaits the full technical report. Where an early figure circulating in secondary coverage could not be corroborated against a primary or independent source, I have left it out rather than repeat it. Nothing in this post is sponsored, and I have no relationship with Moonshot AI.