What Is Kimi K3? The Open 2.8-Trillion-Parameter Model Taking On GPT and Claude

Moonshot AI just dropped Kimi K3, the world’s first open 3T-class model. It packs 2.8 trillion parameters, yet it reportedly activates only 1.8% of them per query to keep inference practical. That’s not just a bigger brain; it’s a new way to scale intelligence without breaking the bank. Here’s what the numbers actually mean for you.

What Is Kimi K3, and Who Built It?

Kimi K3 is a large language model from Moonshot AI. Think of a language model as a giant probability engine: you feed it text, and it predicts what comes next, one token at a time, well enough to write code, answer questions, or hold a conversation. K3 is the current flagship in that family, and it’s massive.

The headline number is 2.8 trillion parameters. Parameters are the adjustable dials a model tunes during training; more of them, in general, means more room to store patterns. But the number that matters just as much is how many of those dials fire on any single question — and that’s where K3 gets clever. A technician adjusting cooling pipes on a custom GPU cluster

Why a 2.8-Trillion-Parameter Open Model Matters

The word doing the heavy lifting here isn’t “2.8 trillion.” It’s “open.” Moonshot is releasing K3 as an open-weight model, which means the actual trained network — not just an API — goes out to the world. Anyone can download it, study it, and fine-tune it.

Until now, models at this scale lived only inside labs like OpenAI, Google, and Anthropic. Handing one to the public is a strategic platform shift. It democratizes frontier-scale intelligence, letting independent researchers, startups, and well-resourced teams build on top of a 3T-class architecture. It’s the difference between renting a race car and being given the blueprints.

Moonshot launched K3 on July 16, 2026, with the full open weights scheduled for release on July 27. The model also features a tiered API pricing structure, making it accessible even if you can’t host it yourself. A cross-functional team gathered around a large touch-screen

How Does It Stack Up Against GPT, Claude, and Gemini?

Strong — with an honest asterisk. Kimi K3 has climbed into the upper tier of frontier text benchmarks, putting it in genuine conversation with the best. Most striking, it claims specialized frontend-coding dominance, beating rivals like Claude Fable 5 at turning instructions into working web interfaces.

That said, K3 does not claim the overall crown. On aggregate benchmarks, it still trails Anthropic’s Fable 5 and OpenAI’s GPT 5.6 Sol. Moonshot says so plainly. The story isn’t “an open model beat everyone.” It’s “an open model got close enough to matter, and won outright in at least one high-value arena.” A developer testing multiple AI applications simultaneously

Hybrid Attention and Sparse MoE, in Plain Terms

K3 introduces two pieces of new plumbing. The first, hybrid attention, is a reworking of the “attention” mechanism — the part of the model that decides which earlier words matter to the word it’s writing now. The second, a sparse Mixture-of-Experts (MoE) design, changes how the model processes tokens.

Instead of running all 2.8 trillion parameters for every token, the model routes each query to a small, specialized slice of itself. According to reports, only about 1.8% of the network activates at a time. This sparsity is what keeps a 2.8-trillion-parameter model from being impossibly slow and expensive to run. It’s a huge brain that only wakes up the neurons it needs.

The Catches

A few, and they’re worth saying out loud: A silicon chip wafer illuminated by bright overhead lighting

Topping a benchmark isn’t the same as being reliable. Leaderboard wins are narrow slices; day-to-day quality across messy real tasks is still being independently evaluated.
You probably can’t run it at home. A 2.8-trillion-parameter model, even sparse and quantized, needs serious hardware. “Open” doesn’t mean “runs on your laptop” — it means the community and well-resourced teams can host and adapt it.
It still trails the closed leaders overall. K3 narrows the gap; it doesn’t erase it.

The real significance of Kimi K3 isn’t a single benchmark. It’s the proof that frontier-scale intelligence can be built, quantized, and given away — and that the open-weight world just moved its ceiling up by a trillion-plus parameters. Whoever leads the leaderboard next month, the blueprints are now in a lot more hands.