Moonshot Releases Kimi K3, the Largest Open Model Yet

Moonshot AI has released Kimi K3, and the headline number is hard to ignore: 2.8 trillion parameters, which the company says makes it the largest open weight model anyone has shipped. It is a sparse mixture of experts model, so each token only wakes up a small slice of that total, and it comes with a one million token context window, native vision, and an always on reasoning mode Moonshot calls thinking mode. Announced on July 16, 2026, with full weights promised for July 27, K3 already lands near the top of a couple of respected benchmarks. For developers weighing whether to route work to a frontier API or eventually self host, this is the release worth reading closely.

Every few weeks a new model arrives claiming a crown, and most of the time the sensible move is to shrug and wait for the dust to settle. Kimi K3 is a little different, and not only because of the eye watering parameter count. An open weight model that trades blows with the closed frontier changes the arithmetic for anyone who has been quietly wondering whether they will always be renting their intelligence from someone else's API.

The number that matters, and the one that does not

Two point eight trillion parameters is the figure everyone will repeat, and it is genuinely large. But the honest way to read a mixture of experts model is to look at what actually runs. K3 routes each token through just 16 of its 896 routed experts, so the model behaves, for the cost of a single forward pass, far more like a mid sized model than its headline total suggests. That is the whole point of the sparse design: you get the knowledge capacity of a giant while paying compute closer to something you can actually serve.

Underneath sit two pieces of Moonshot's own engineering. Kimi Delta Attention is a hybrid linear attention mechanism that keeps the memory cost of a very long context in check. Attention Residuals is what the company describes as a drop in replacement for the usual residual connections, one that it says gives steadier gains as the model scales. You do not need to love the internals to care about the result, which is a one million token context window that stays affordable enough to use rather than just advertise.

Where it lands against the frontier

Benchmarks are noisy and easy to overfit, so read these as a signal rather than a scoreboard. On GDPval-AA v2, which scores real tasks across 44 occupations and 9 industries, K3 came in at 1,687. That put it third overall, behind Claude Fable 5 Max at 1,815 and GPT-5.6 Sol Max at 1,747.8, and comfortably ahead of Claude Opus 4.8 at 1,600. On the Frontend Code arena, where developers blind test the models on building interfaces, K3 finished first at 1,679, nudging past Fable 5.

The one line summary making the rounds is that K3 offers roughly Opus class quality at a fraction of the price. That is a marketing gloss, not a law of nature, and your own evals on your own workloads will tell you more than any leaderboard. Still, a top three frontier result from a model whose weights you can download in a week is the kind of thing that reshapes buy versus build conversations.

Self hosting is a server job, not a laptop one

Here is the reality check that the parameter count implies. You are not running full K3 on a workstation. Moonshot recommends at least 64 accelerators to serve the model, and even an aggressively quantized build needs far more memory than any single card carries. For reference, the lab's earlier one trillion parameter K2.7 Code needs roughly 577 gigabytes of VRAM at four bit precision, and K3 is nearly three times larger. So plan for a multi GPU server, a rented cluster, or a quantized deployment on serious hardware.

The encouraging part is that Moonshot is doing the plumbing so the community does not have to reverse engineer it. The lab is contributing a vLLM prefill cache implementation for its Kimi Delta Attention, timed to release alongside the weights on July 27. That means a serving command as ordinary as the one below should work on day one, rather than months later once tooling catches up.

If you want open weights you can actually run today while you wait, Moonshot's own K2.7 Code remains the practical pick for local coding, and it slots into the same vLLM and SGLang workflows. Treat K3 as the frontier option you reach for through the API now and self host later, once you have the hardware and the weights are public.

What to do with this

For most teams the move this week is not to rip anything out. It is to run K3 through the API against your real prompts and see whether it holds up on the work you actually do, especially long context tasks and frontend code, which is where its published numbers are strongest. The cache hit pricing rewards agents that reuse a big shared system prompt, so if that describes your setup, the cost math may look better than the headline output price suggests.

The longer arc is the one worth watching. An open weight model landing in the top tier means the gap between what you can rent and what you can own keeps narrowing. That does not make the closed labs irrelevant, but it does give you leverage, a fallback, and a path to keeping sensitive workloads on hardware you control. On July 27, when the weights go public, that path gets a lot shorter.

Sources and further reading

• VentureBeat: China's Moonshot AI releases Kimi K3, the largest open source model ever
• MarkTechPost: Moonshot AI releases Kimi K3, a 2.8 trillion parameter open MoE model
• Tom's Hardware: Moonshot releases 2.8 trillion parameter Kimi K3
• Simon Willison: Kimi K3, and what we can still learn from the pelican benchmark
• Latent Space: Kimi K3 2.8T-A50B, the largest open model ever released

Frequently asked questions