Kimi K3 and the open-weight frontier

On July 16, Moonshot AI released Kimi K3. It is a 2.8-trillion-parameter open-weight model, the largest ever released. On GDPval-AA v2, a benchmark measuring real-world tasks across 44 occupations, it scored 1,687, third overall, behind Claude Fable 5 Max and GPT-5.6 Sol Max. It set a state-of-the-art 91.2 on BrowseComp, a benchmark for long-horizon information seeking. The full weights ship July 27.

Two months ago we wrote about GLM-5.2 landing within a point of the frontier on coding benchmarks. K3 is the next data point on the same line. The gap between open and closed is not closing asymptotically. It is closing in steps, and the steps are getting larger.

What the model is

K3 is a sparse Mixture-of-Experts model: 2.8 trillion total parameters, 16 of 896 experts active per token. It uses two architectural innovations from Moonshot: Kimi Delta Attention, a hybrid linear attention mechanism, and Attention Residuals, a drop-in replacement for residual connections. It has a 1-million-token context window and native vision. Training used quantization-aware MXFP4 weights with MXFP8 activations from the SFT stage onward, for broad hardware compatibility.

The scale is the headline, but the efficiency argument is more interesting. Moonshot reports roughly 2.5× improvement in scaling efficiency over K2, more intelligence per unit of compute. The architecture is designed to scale well beyond the trillion-parameter regime. This is not a one-off stunt; it is a trajectory.

The benchmark picture

On coding, K3 trades blows with the frontier. It leads on SWE Marathon and Program Bench. On Terminal-Bench 2.1 it trails GPT-5.6 Sol by half a point. On FrontierSWE it is competitive with Claude Fable 5. On agentic benchmarks (AutomationBench, SpreadsheetBench 2, BrowseComp), it ranked first in four of eight.

Moonshot's own 48-hour chip design demo is the most revealing signal. Given an open-ended hardware engineering task and no human intervention, K3 completed a full chip construction pipeline (architectural design, optimization, verification) using open-source EDA tools on a 45nm library. The result was a 4 mm² chip that closes timing at 100 MHz and decodes over 8,700 tokens per second in simulation: a chip designed by a model, for a model. Whether that is impressive or alarming depends on where you sit.

The sovereignty angle, again

GLM-5.2 made the argument two months ago: when an open model lands near the frontier, the premium you pay for a closed API is for convenience, not quality. K3 makes it again, louder. The API is priced at $3 per million input tokens and $15 per million output. Cached input drops to $0.30. But the API is the convenience tier. The weights are the point.

A 2.8-trillion-parameter model you serve yourself does not need a terms-of-service page, because it does not have one.

The geopolitical dimension is impossible to ignore. US officials and Anthropic have accused Moonshot of distilling American models, training on outputs from closed systems. Moonshot is backed by Alibaba. The release was timed to land ahead of the World AI Conference in Shanghai. Open-sourcing the world's largest model is, among other things, a bid to become the center of gravity for the global open-source developer community. That is a strategy, not an accident.

For enterprises, the question is not whether to trust Moonshot. It is whether the model you serve yourself, on hardware you control, behind a firewall you configure, is more or less trustworthy than a vendor API whose pricing, retention, and availability can change on someone else's schedule. The model takes the same prompt either way. The data path is the part that differs.

The infrastructure reality

None of this is free. A 2.8-trillion-parameter model does not run on a single server rack. Moonshot recommends deploying K3 on supernode configurations with 64 or more accelerators. Inference at this scale requires the disaggregated serving architecture Moonshot pioneered with Mooncake (separating prefill from decode, with KV-cache-centric scheduling) to make the economics work at all.

This is the part vendors selling sovereignty tend to skip. The model is open, but the hardware is not. Running K3 well demands real compute, real networking, and real operations. That is the trade, and it has not changed since we last wrote about it.

K3 on PrivateMind

Options Technology is evaluating K3 now. When the weights land, K3 will be available on the PrivateMind platform behind the same gateway, with the same audit trail, on the same infrastructure as every other model we serve. No prompt leaves. No token goes to a third party. The model is 2.8 trillion parameters. The data path has zero external dependencies.