2.8 trillion. The open-weights ceiling just broke again.
At midnight on July 17, Moonshot AI dropped Kimi K3.
2.8 trillion parameters, Mixture-of-Experts, 16 experts fired per token — the first open model to cross the 2-trillion threshold. By that measure alone, it's a milestone.
The headline: in their own benchmarks, it ranks just behind Claude Fable 5 and GPT-5.6 Sol on overall intelligence. For the first time, an open model is breathing down the neck of the closed frontier.
It's not just scale
Moonshot didn't just pour in parameters. K3 runs on two homegrown architectural tricks:
- Kimi Delta Attention (KDA) — keeps information flowing smoothly across very long sequences
- Attention Residuals (AttnRes) — cross-layer residual connections, the deeper it goes the steadier it stays
- Stable LatentMoE — 16 of 896 experts activated, roughly 2.5x better scaling efficiency than K2
In plain terms: parameters went up, but inference cost didn't climb in lockstep. That's the only way open models can actually compete with closed ones. Big and expensive defeats the purpose.
The coding numbers hold up
A few hard figures:
- DeepSWE: 67.5%
- Program Bench: 77.8%
- SWE Marathon: 42.0%
Software engineering is where K3 flexes hardest. It reads source code, watches render results, parses run logs, and decides the next fix from test feedback — not a glorified autocomplete, more like an engineer you can hand a ticket to.
Wilddest demo: K3 autonomously designed a chip on a 45nm process in a single 48-hour run — building, optimizing, and verifying it with open-source EDA tools. As a side quest, it also wrote MiniTriton, a compact GPU compiler from scratch, rivaling Triton and torch.compile.
Pricing and the open timeline
API pricing (per million tokens):
- Cache hit: $0.30
- Cache miss input: $3.00
- Output: $15.00
$15 output isn't cheap for an open Chinese model — the old playbook was rock-bottom pricing. The posture shifted this time: not racing to the bottom, pricing on capability instead. Full weights drop July 27.
The company behind it
Moonshot has closed 6 rounds this year, latest pre-money valuation $31.5 billion, annual recurring revenue past $300 million, with API revenue north of 70% of that.
There's a clean read on this: classic "commoditize your complement" — drive the cost of intelligence toward zero, make money on the infrastructure underneath. The Chinese open-weights labs are moving faster and faster.
A breath of cold air
A few caveats worth flagging:
- 2.8T is total params — activated is around 41B class. Don't conflate total params with inference cost
- Official benchmarks are official. Independent third-party runs (Artificial Analysis and the like) have to wait until the weights actually land
- $300M ARR is mostly API, which means developers are genuinely using it — but enterprise-grade deployment stories are still thin
Still, this is the first time an open model has pushed scale and efficiency to this level at once. July 27, when the weights land, is the real exam.