AI Models

Kimi K3 Benchmarks: Is China's 2.8T Open Model Really Beating GPT-5.6?

Newaiera Desk · 2026-08-03 · 7 min read

Moonshot AI released the largest open-weight model ever built and it topped two coding leaderboards. We went through the actual scores — the gaps are small

The headline everyone ran was that a Chinese open-weight model beat GPT-5.6. The scores say something more interesting, and more useful if you are actually choosing a model to build on.

On July 16, Beijing-based Moonshot AI released Kimi K3, a 2.8-trillion-parameter mixture-of-experts model. It is the largest open-weight model anyone has published. Moonshot bills it as the first open "3T-class" system, and it landed at or near the top of several coding leaderboards within days.

So: is it actually beating the frontier labs? On two benchmarks, narrowly, yes. On the rest, it is close enough that the difference stops being the point.

The actual coding scores

Here is where K3 landed against GPT-5.6 Sol on the three coding benchmarks with published head-to-head numbers.

Two firsts and one near-miss. But look at the margins. On Program Bench the gap is two tenths of a point. On Terminal Bench it is half a point the other way. Those are not differences you would notice in a working day; they are inside the noise band you get from re-running most agentic benchmarks with a different random seed.

SWE Marathon is the one worth taking seriously. It runs 20 realistic tasks over multi-hour sessions with hidden checks, network restrictions and exploit scans — a much better proxy for "can this thing be left alone with a repository" than a single-shot coding test. A two-point lead there is a real result.

The architecture is the trick

2.8 trillion parameters sounds unaffordable to run. It is not, because almost none of them fire at once.

K3 is a mixture-of-experts model with roughly 104 billion active parameters per token. The other ~96% of the network sits idle on any given forward pass. That is why a model with a headline parameter count four to five times larger than most frontier systems can serve tokens at a price that is not absurd.

The context window is 1,048,576 tokens — a genuine million-token window, not a marketing round-up. Moonshot also cites KDA and attention-residual changes for efficiency, though the independent verification on those is thin so far.

Pricing is the actual story

This is the part that will decide whether K3 matters commercially.

At $2.90 in and $14.00 out per million tokens, K3 undercuts the frontier closed models on the workloads where it is competitive. Moonshot claims typical customer costs land 60–80% below list once prompt caching is in play, which is plausible for agentic work — long-horizon coding sessions re-send enormous amounts of unchanged context, and caching is exactly what that pattern rewards.

Pair that with open weights and the calculus changes for anyone with data-residency constraints. You are no longer choosing between "frontier quality" and "runs where I need it to run."

What to actually take from this

Three honest conclusions.

The open/closed quality gap on coding is effectively gone at the top end. Not "narrowing" — gone, within measurement error, on the benchmarks that have been published. That was not true twelve months ago.

Benchmark wins of 0.2 points are marketing, not engineering. If you are picking a model, the Program Bench result should not move you. Latency under your real prompt shape, tool-calling reliability, and how the thing behaves in hour three of an agent run will matter far more than any of these numbers.

Compute restrictions are not working the way they were meant to. Moonshot built this under US export controls on advanced accelerators. Whatever you think the policy was supposed to achieve, a 2.8T open-weight model at the top of the coding leaderboards is not it.

The caveat worth stating plainly

All of the above rests on published benchmark results, several of them reported by the labs themselves or by secondary coverage rather than independent evaluators. Coding benchmarks are also the most contaminated category in the field — training sets and evaluation sets overlap more than anyone would like, and a model trained more recently has seen more of the internet's solutions.

Treat the ranking as a signal, not a verdict. The pricing and the open weights are the facts that hold up regardless of whose leaderboard you believe.

A model that is 0.2 points behind and 70% cheaper is not losing.

If you are running agentic coding workloads at volume, K3 is worth a genuine evaluation against your own tasks. If you are picking a model because a leaderboard put it first, you are optimising the wrong thing.