Kimi K3: Moonshot’s 2.8 Trillion Parameter Model Hides a Hallucination Problem

Moonshot AI’s Kimi K3 is the largest open-weight language model ever released, and starting July 27, 2026, anyone can download all 2.8 trillion parameters of it — but an independent lab found its hallucination rate jumped to roughly 51%, a number that never made it into Moonshot’s own benchmark charts.

I’ve been tracking open-weight releases pretty closely this year because the gap between “open” and “frontier” keeps shrinking, and Kimi K3 is the clearest sign yet of where that’s heading. This isn’t a scaled-down community model somebody trained in a dorm room. It’s a 2.8-trillion-parameter system from a well-funded Chinese lab, and on paper it beats or matches several models that cost a fortune to access through an API. That combination — frontier-class performance, zero-dollar weights — is exactly the kind of thing that reshapes who gets to build on top of AI, for better and worse.

What Kimi K3 actually is

Kimi K3 comes from Moonshot AI, the Beijing-based lab behind the earlier Kimi K-series models. It’s a sparse mixture-of-experts (MoE) architecture, meaning the full model has 2.8 trillion parameters but only a small slice of them fire for any given token — in K3’s case, 16 out of 896 experts get activated per token. That’s the trick that makes a model this size usable at all: you get the knowledge capacity of a 2.8T-parameter network without paying the compute cost of running all of it every time.

The model shipped initially on July 16, 2026, with API access, and Moonshot has now locked in July 27 as the date the full weights land on Hugging Face for anyone to download and run themselves. Alongside the raw scale, K3 ships with a 1-million-token context window, native vision support (it can read images, not just text), and what Moonshot calls an “always-on” thinking mode, where the model reasons step by step before answering rather than that being an optional toggle.

The benchmark numbers, and where they get complicated

On GDPval-AA v2, a benchmark meant to approximate real economically valuable work rather than trivia-style Q&A, Kimi K3 scored 1,687 — putting it third overall, behind Claude Fable 5 Max and GPT-5.6 Sol Max, but ahead of Claude Opus 4.8. That’s a genuinely notable result for an open-weight model to post against the best closed frontier systems money can buy.

K3 also took the top spot in the Frontend Code Arena with 1,679 points, edging out Claude Fable 5 (1,631), GPT-5.6 Sol (1,618), and GLM-5.2 (1,587). If you build web interfaces and want a model you can self-host without an API bill, that’s the kind of number that gets attention fast.

But here’s the part that made me actually sit up: independent testing by Artificial Analysis found K3’s hallucination rate climbed to roughly 51%, up from about 39% on its predecessor model. That’s not a rounding error — that’s the model being wrong, confidently, on more than half of whatever this eval measured. And it’s a number you will not find anywhere in Moonshot’s own published benchmark charts. Whether that’s an oversight or a choice, the effect is the same: the marketing numbers look great, and the reliability number that actually matters if you’re putting this in front of users doesn’t show up until someone outside the company goes looking for it.

Why the storage number matters as much as the parameter count

2.8 trillion parameters sounds abstract until you look at what it costs to actually have the thing sitting on a drive. Moonshot ships K3 in MXFP4, a 4-bit weight format, which brings the full download down to about 1.4 terabytes. Compare that to the roughly 5.6 terabytes the same parameter count would need at full FP16 precision, and you can see why the quantization choice isn’t optional — it’s the difference between “a serious GPU cluster can host this” and “nobody but a handful of labs could ever touch it.”

Even at 1.4TB, though, this isn’t a model you’re running on a gaming PC. Self-hosting K3 realistically means a multi-GPU server setup, and that’s before you get into the inference economics of serving it to actual users at any kind of scale. “Open weights” here means open to organizations with serious infrastructure, not open to a hobbyist with a single RTX card.

The bigger pattern: open-weight models are closing the gap fast

K3 isn’t happening in isolation. DeepSeek’s V4 model reached general availability on July 20 and fully retired its older API aliases on July 24 — the same week Moonshot has been building up to K3’s open-weight drop. Between the two releases, this is arguably the single biggest concentration of large open-weight launches in one month all year. If you’ve been assuming “open-weight” means “a step or two behind the frontier,” that assumption is aging badly in real time.

That matters beyond the leaderboard bragging rights. When frontier-tier capability becomes downloadable rather than API-gated, it changes who has access to it, how startups build products, and also who’s responsible when something goes wrong. A closed model’s provider can push a fix or pull an API key. A downloaded 2.8-trillion-parameter model with a 51% hallucination rate on some tasks is out in the world, running wherever anyone chooses to run it, unpatched until the next release.

What this means if you’re actually deciding whether to use it

If you’re a developer or a company evaluating K3 for a real project, the coding benchmark numbers are worth taking seriously — a top spot in the Frontend Code Arena isn’t nothing. But the hallucination finding is the number I’d want answered before shipping anything user-facing on it. A near-frontier coding model that’s also confidently wrong roughly half the time on some evaluation set is a very different tool depending on whether your use case is “generate boilerplate a human will review” or “answer questions directly to end users.” The first is probably fine. The second needs a lot more caution than the marketing slides are going to tell you.

This whole pattern — big capability numbers up front, reliability concerns you only find by reading the independent testing — isn’t unique to Moonshot. It’s part of why groups like the Future of Life Institute have started grading AI labs on safety practices rather than just capability, and why no lab scored above a C+ on the most recent safety index. Capability keeps racing ahead of the evaluation and disclosure practices meant to keep pace with it.

FAQ

Is Kimi K3 free to use?

The weights themselves are open and free to download starting July 27, 2026, but running a 2.8-trillion-parameter model requires serious multi-GPU hardware. Moonshot also offers paid API access if you don’t want to self-host.

How does Kimi K3 compare to Gemini, GPT, and Claude models?

On the GDPval-AA v2 benchmark, K3 ranks third behind Claude Fable 5 Max and GPT-5.6 Sol Max, ahead of Claude Opus 4.8. It leads the Frontend Code Arena outright. For a broader look at how the current frontier models stack up, see our breakdown of Google’s latest Gemini releases.

What does “open-weight” actually mean here?

It means Moonshot is publishing the trained model parameters themselves (in MXFP4 format, about 1.4TB total) rather than only offering access through a paid API. Anyone with the hardware can download and run the model without Moonshot’s involvement.

Why is the hallucination rate finding significant?

Because it wasn’t disclosed by Moonshot in its own benchmark materials — it came from independent testing by Artificial Analysis, which found the rate climbed from about 39% on the previous K-series model to roughly 51% on K3. That’s a meaningful reliability regression for a model being marketed primarily on its capability gains.