Kimi K3 — Moonshot swings for the largest open model ever, API first

Moonshot AI launched Kimi K3 on July 16 — a 2.8-trillion-parameter Mixture-of-Experts flagship (~50B active per token) with a 1M-token context window, native vision, and a new hybrid-attention architecture (Kimi Delta Attention plus Attention Residuals). It is API-first: live on Kimi.com, Kimi Code, and api.moonshot.ai at $3 input / $15 output per 1M, with open weights promised by July 27. If the weights ship, it would be the largest open-weight model ever released. It still changes no local pick — 2.8T is datacenter-class under any quantization — but this is the marquee drop of the window.

Verdict: The biggest open-weights promise ever made — 2.8T parameters, 1M context, Opus-class scores at Sonnet-class pricing — but the weights aren't out yet and never fit a home rig anyway

The take

The facts, verified against Moonshot's own platform docs (platform.moonshot.ai — `kimi-k3`, $3.00 per 1M input on cache miss, $0.30 on cache hit, $15.00 output, context 1,048,576 tokens) and the official announcement: 2.8T total parameters, roughly 50B active per token, 1M context, native multimodal input, always-on reasoning (the `reasoning_effort` field currently supports only `max` — you cannot turn thinking off, so the output-token bill runs structurally high at $15/1M). Weights are NOT downloadable yet: Moonshot's announcement promises open weights by July 27, the same open-frontier playbook the K2 line followed. Until a license and model card land on Hugging Face, treat "open" as a promise, not a property.

Where it sits: Artificial Analysis scored K3 at 57 on its Intelligence Index — comparable to Opus 4.8 and GPT-5.5, behind Claude Fable 5 and GPT-5.6 Sol. Moonshot's own numbers (93.5% GPQA Diamond, 88.3% Terminal-Bench 2.1, 91.2% BrowseComp) are vendor-reported and lean agentic — long-horizon coding and tool use is the positioning. The pricing is the aggressive part: $3/$15 is Sonnet-5-standard-tier money for near-Opus capability, roughly 10× under Fable 5 on input. The economics caveat is the always-on thinking: your effective per-task cost depends heavily on how many reasoning tokens it burns, which you cannot cap at launch.

Our call: no planner-pick change, and no model entry until the weights actually land. Even when they do, 2.8T total (~1.4 TB at 8-bit, still hundreds of GB at aggressive quants) is beyond every box on our hardware list by an order of magnitude — this joins LongCat-2.0, Hy3, and GLM-5.2 in the hosted-frontier-comparator bucket, and would take the "largest open weights ever" crown from LongCat's 1.78T. What we watch for on July 27: the actual license (K2 was Modified MIT), the model card, and whether the checkpoint that ships matches the hosted one. The pattern to appreciate: the open-weight frontier is now within one step of the closed frontier — and it keeps being Chinese labs closing the gap.

Where this fits

Models: Kimi K2.6 · GLM-5.1 · DeepSeek V4-Pro · Qwen 3.6-35B-A3B

Hardware: NVIDIA DGX Spark · Dual RTX 5090 · Mac Studio M3 Ultra 96 GB

Sources

Next step

Try this in the planner→