Moonshot AI shipped Kimi K3 on July 16, and nearly every headline led with the same number: 2.8 trillion parameters, the largest open-weight model announced to date. That number is real, and it is the least informative thing about the release. K3 is a sparse Mixture-of-Experts model that fires 16 of its 896 experts per token — one token sees about one fifty-sixth of the expert pool — so the 2.8 trillion is mostly latent capacity that never runs on any given forward pass. The figure that will decide whether this model matters is not its size but its efficiency: a hybrid linear-attention design, a claimed 2.5x better scaling efficiency than Kimi K2, 21% fewer output tokens than K2.6 on equal tasks, MXFP4 quantization, and roughly half the per-task cost of Claude Opus 4.8. If those hold up under the July 27 technical report, the story is that the frontier is increasingly being won on architecture and algorithms, not raw scale.
I want to make the reframe precise, because the coverage keeps committing the same category error, and the error is not harmless — it tells founders to evaluate models on the one axis that means the least.
2.8 trillion is almost a vanity metric
In a dense transformer, every parameter participates in every token. The parameter count is then a fair proxy for the compute each token costs, and comparing two dense models by their size tells you something. A sparse MoE breaks that proxy on purpose. The model holds a large bank of experts — here, 896 of them — and a router selects a small subset for each token. K3 activates 16. The other 880 sit idle for that token and cost nothing on that pass.
So the arithmetic that matters is not 2.8 trillion. It is 16 of 896, which is 1.8% of the expert pool, or one expert set in fifty-six. Per-token compute scales with the experts that actually fire, not the ones that exist to be routed to. The 2.8 trillion is a measure of how much specialized capacity the model can store and select from, which is a genuine design lever, but it is not a measure of what the model runs when it answers you.
The announcement gives the total and the routing but, as of writing, does not publish an explicit active-parameter count — that is one of the things the full report should settle on July 27. What the routing alone establishes is that quoting 2.8 trillion against a dense model's parameter count, or even against another MoE's total, compares a warehouse to a checkout lane. Both hold inventory. Only one of them is the throughput you pay for.
This is why I call the headline number close to a vanity metric for MoE models. It is the easy number to print and the hardest one to reason from. When a release brags about total parameters, it is telling you about the size of the store, not the cost or quality of a purchase — and coverage that treats the two as interchangeable is quietly training its readers to buy the wrong thing.
The efficiency stack is what actually moved
Strip the size headline away and look at what Moonshot claims it engineered. Two innovations carry the announcement, and both are efficiency plays.
The first is Kimi Delta Attention (KDA), described as a hybrid linear attention mechanism, with a reported speedup of up to 6.3x for decoding in million-token contexts. This is the one that matters most for anyone doing long-context work. Standard attention cost grows quadratically with sequence length — double the context, quadruple the attention compute — which is exactly why long windows are expensive and slow at the tail. Linear-attention variants trade some of that exact recall for a cost that grows closer to linearly. At K3's 1-million-token window, the quadratic term is the binding constraint, so a hybrid design that blunts it is not a nicety. A 6.3x decode speedup at the top of the context range is the difference between a 1M window being a spec-sheet number and a usable one.
The second is Attention Residuals (AttnRes), which Moonshot says "selectively retrieves representations across depth rather than accumulating them uniformly," for roughly 25% higher training efficiency at under 2% additional cost. Do the ratio: a 25% gain bought for a 2% overhead is better than a twelve-to-one return on the added compute. Training efficiency compounds across the whole run, so a trade like that, if real, shows up as either a cheaper model at the same quality or a better model at the same budget. My read: that is the lever that lets a lab outside the largest compute clusters stay in the race.
Around those sit the rest of the efficiency stack, all Moonshot self-reported: a claimed 2.5x better overall scaling efficiency than Kimi K2; 21% fewer output tokens than K2.6 on equivalent tasks; and MXFP4 weights with MXFP8 activations — four-bit weights, which cut memory and bandwidth against the usual eight- or sixteen-bit baselines. The other named components (Stable LatentMoE with Quantile Balancing, Per-Head Muon, Sigmoid Tanh Unit, Gated MLA) are, without the paper, names on a list; I will not pretend to evaluate them before July 27.
The token-efficiency number deserves a second look because it converts directly into money. K3's output is priced at $15.00 per million tokens against $3.00 for input, so output is the expensive side of any reasoning or coding workload. Emitting 21% fewer output tokens on the same task is a 21% cut to the dominant cost line — before you count the headline rate. Stack that on the pricing and the "half the per-task cost of Opus 4.8" claim stops being a slogan and starts being an arithmetic consequence of shorter answers plus a lower rate.
Every item in this section is efficiency-per-capability. None of it is size. That is the release.
Sort the benchmarks by who is vouching
The scores are strong, but they come from two different kinds of source, and the distinction is the whole game.
Third-party. On Artificial Analysis's composite leaderboard, K3 posts an Elo of 1,547 — behind only Claude Fable 5 and ahead of GPT-5.5 — which is a 732-point jump from Kimi K2.6. On Arena.AI's Frontend Code Arena, K3 sits at #1 with 1,679, ahead of both Claude Fable 5 and GPT-5.6 Sol. Those are independent measurements, and they are the ones I weight.
Moonshot self-reported. The company says K3 "mostly beats Claude Opus 4.8 max and GPT-5.5 high" while trailing Claude Fable 5 and GPT-5.6 Sol, with reported leads on Program Bench (77.8), SWE Marathon (42.0), BrowseComp (91.2), Automation Bench (30.8), and OmniDocBench (91.1). Plausible, consistent with the third-party picture, and not yet independently confirmed.
Hold the two results side by side and note the nuance, because it is easy to misread in either direction. K3 is #1 on the frontend-code board and #2 on the broader composite, behind Fable 5. Both are true. They are different benchmarks measuring different things — one a specific coding surface, the other a general blend — so a model can top the narrow one and place second on the wide one without contradiction. The honest summary is not "the new #1 open model" and not "second place." It is: best-in-class on frontend code, second on the general composite, at a fraction of the price.
| Metric | Kimi K3 | Source |
|---|---|---|
| Artificial Analysis composite Elo | 1,547 (behind Fable 5, ahead of GPT-5.5) | Third-party |
| Frontend Code Arena | #1, score 1,679 | Third-party (Arena.AI) |
| vs Opus 4.8 max / GPT-5.5 high | "mostly beats" | Self-reported |
| Input / output price (per M tokens) | $3.00 / $15.00 | Moonshot |
| Cache-hit input | $0.30 | Moonshot |
| Per-task cost vs Opus 4.8 | ~half | Moonshot |
The pricing has one more edge worth naming. Cache-hit input runs $0.30 against the $3.00 cache-miss rate — ten times cheaper on repeated context — and Moonshot reports above 90% cache hits in coding workloads. For agentic coding, where the same repository context is re-sent turn after turn, that is where the real bill is decided.
Hold the chip demo at arm's length
The most cinematic claim in the materials is the one to trust least for now. Moonshot reports that K3, tasked with designing a physical chip to run a nano-scale version of itself, completed the full pipeline — architectural design, optimization, verification — over 48 hours of continuous autonomous operation using open-source EDA tools. It is a striking demonstration and a good story.
It is also a company demonstration that has not been independently verified, and self-designing hardware is exactly the kind of result where the gap between "ran once in a curated setup" and "reproducible capability" tends to be widest. I treat it as a marketing artifact until a third party reproduces it. The load-bearing evidence in this release is the efficiency stack and the third-party benchmarks, not the chip.
A bet the evidence supports, not a settled fact
Everything above rests on a schedule. As of roughly July 18, the weights are not public and the complete technical report is not out; both are due July 27 on Moonshot's Hugging Face org under a modified MIT license. Until then, the architecture comes from an announcement and preliminary materials, not a peer-reviewable paper, and the efficiency figures are mostly self-reported. I am arguing that the direction of the evidence is clear, not that the numbers are final. Watch what July 27 does to them.
The reframe matters because of where the frontier's headwinds actually are. Scaling was always a curve without a mechanism — an extraordinary empirical regularity and an explanation of nothing — and the returns to raw parameter count are precisely the place the curve is bending. If the marginal capability bought by another order of magnitude of dense scale is compressing, then the lab that wins is the one that lowers the cost of a unit of capability. That is the entire content of K3's efficiency stack: KDA on decode, AttnRes on training, token efficiency on output, MXFP4 on memory. The release is legible only if you read it as a cost-of-capability move.
There is a second reading that K3 makes almost against its own headline. The market reaction — stocks moving, commentators reaching for "a second DeepSeek shock" — was priced off the size and the price gap, an open-weight model closing on the closed US frontier at a fraction of the cost. Fair enough. But the durable lesson runs the other way, toward small, well-engineered, composable systems over the biggest available model. K3 is enormous, and yet its advances are the same efficiency and orchestration advances that make bounded, sparsely-activated systems attractive in the first place. Sparse routing is composability inside one set of weights: most of the model stays dark, and only the relevant experts light up. Even the largest open model now competes on how little of itself it has to run.
So the actionable version, for anyone who evaluates models for a living: track efficiency per unit of capability, not parameter counts. Cost per solved task. Output tokens per task. Decode throughput at your real context length. Elo per dollar on a benchmark that resembles your workload. Those numbers tell you what a model costs to be useful. The parameter count tells you how big the warehouse is, and for a Mixture-of-Experts, that is a number the routing was built to make you stop caring about.
The largest open model yet announced, and the most important thing about it is how little of itself it uses.