Kimi K3: Moonshot's 2.8T Open Model Launches

Moonshot AI released Kimi K3 on July 16, 2026: a 2.8-trillion-parameter open-weights model with a 1-million-token context window, native multimodal input, and a new attention scheme the company calls Kimi Delta Attention. The API is live now. The weights are scheduled to go public on July 27, per Moonshot's launch materials and reporting from the FT, TechCrunch, and Reuters.

That two-step rollout, paid API first and open weights eleven days later, is the shape of the whole announcement. You can build on K3 today through Moonshot's endpoints. If you want to run it yourself, fine-tune it, or audit it, you wait until the 27th.

2.8T parameters now, open weights on July 27

The headline number is 2.8 trillion total parameters. That would make K3 the largest model Moonshot has published and one of the largest open-weights releases anyone has shipped. Meituan's LongCat-2.0, which topped coding charts earlier this month, ran 1.6T. K3 nearly doubles it.

A caveat worth stating plainly. A 2.8T total-parameter count almost certainly describes a mixture-of-experts model, where only a fraction of those weights fire per token. That is how everyone builds at this scale, and Moonshot's own Kimi K2 was a sparse MoE. But as of launch I have not seen Moonshot publish the active-parameter count for K3, which is the number that actually sets inference cost and hardware footprint. Until that arrives with the weights on the 27th, read "2.8 trillion" as a capacity figure, not a serving-cost figure.

The 1M context window puts K3 in the same bracket as the closed frontier on paper. GPT-5.6 Sol and Google's Gemini line both sit at or above that mark. What context length doesn't tell you is retrieval quality across the window, and Moonshot has not released long-context recall numbers. A million tokens you can't reliably attend to is a spec-sheet flex, not a feature.

Kimi Delta Attention is the line to read twice

The most interesting technical claim in the release isn't the parameter count. It's Kimi Delta Attention, Moonshot's name for the attention mechanism K3 uses in place of standard full attention.

Moonshot is positioning it as the reason a 2.8T model with a 1M window is economically serveable at all. The pitch, as I read it, is a linear or near-linear attention variant that keeps memory and compute from exploding as context grows. If that holds up, it's the difference between a model that can technically accept a million tokens and one you can afford to actually feed a million tokens on a regular basis.

The honest position here is skepticism until the weights land. Efficient-attention claims are common and the failure mode is always the same: the mechanism saves compute but quietly degrades quality on the long-range dependencies that were the entire point of the big window. Moonshot's advantage is that it's shipping the weights, so within days of July 27 independent researchers will be able to test exactly that. Few closed labs give you the option.

This is K2's sequel, and the strategy hasn't changed

Kimi K2, released in 2025, was Moonshot's first real dent in the open-weights conversation: a large MoE that punched above the attention it got in Western coverage. K3 is the same play, scaled up and dressed with multimodal input and the new attention stack.

The business context matters. Moonshot is reportedly raising at a $30 billion valuation, a number that only makes sense if the company is seen as a genuine frontier lab rather than a fast-follower. Shipping a 2.8T open model with a firm weights date is exactly the kind of move that supports that story to investors. Open weights are a distribution strategy as much as a research one. They put Kimi in front of every developer who won't pay for an API but will happily download a checkpoint, and they build the mindshare that a Chinese lab can't easily buy through Western channels.

The Chinese open-weights bench is getting crowded

K3 doesn't arrive into empty space. It joins a run of Chinese open releases that have been landing almost weekly.

• DeepSeek made its V4-Pro price cut permanent and remains the reference point for cheap, capable open models.
• GLM-5.2 from Z.ai topped several coding benchmarks on release.
• Meituan's LongCat-2.0 shipped a 1.6T open model that led coding charts earlier in July.

Set against that group, K3's differentiators are raw scale, the native multimodal input, and Kimi Delta Attention. What's conspicuously missing is a benchmark table putting K3 against those exact rivals. Moonshot is competing in the most benchmark-saturated segment in AI and launched without the head-to-head numbers everyone in it publishes by default. That's a choice, and it's a strange one for a model this large.

Cheap open weights squeeze the closed labs

The reason a US developer should care about a Chinese model they may never deploy is pricing pressure. Every credible open frontier release resets what buyers expect to pay for closed-model tokens.

When K3's weights are public on July 27, inference providers can host them, and hosted open models tend to undercut first-party closed APIs by a wide margin. That's the same dynamic DeepSeek forced. A team currently paying Sol or Claude Fable 5 rates for a workload that a hosted K3 could handle now has a live alternative to price against, even if they'd rather not switch. The closed labs keep their edge on the hardest reasoning and agentic tasks, where benchmarks and real usage still favor GPT-5.6 and Fable 5. But the broad middle of the market, the summarization, extraction, and long-document work that a big-context open model does fine, is exactly where open weights erode margin.

All figures from each vendor's own release materials; the dashes mark specs I couldn't confirm from a primary source, not zero.

The benchmark table is what Moonshot didn't ship

Three things decide whether K3 is a genuine frontier entry or a very large model that reads well in a press release, and none of them are answerable today.

The active-parameter count sets whether K3 is realistically serveable outside a well-funded lab. Independent long-context recall numbers determine whether that 1M window works or just accepts input. And Kimi Delta Attention needs third-party evaluation before anyone should believe it delivers efficiency without a quality tax.

The good news is that Moonshot committed to a date. On July 27 the weights become testable, and every one of these open questions moves from Moonshot's claims to reproducible results. That's the argument for open weights in one sentence: you don't have to trust the launch post, you just have to wait eleven days and run the model yourself.