AI Model Report

Open Source · JULY 17, 2026

Moonshot ships Kimi K3 at 2.8T parameters, third on Intelligence Index

Beijing's Moonshot AI released Kimi K3 on July 16 — a 2.8-trillion-parameter sparse MoE with Kimi Delta Attention, a 1M-token context window, and API pricing at $3/$15 per million tokens. Full weights are due July 27.

By Lars Iverson · Open source & model weights · July 17, 2026

Moonshot AI released Kimi K3 on July 16, a 2.8-trillion-parameter sparse mixture-of-experts model that Beijing's most closely watched lab claims sits within striking distance of Claude Fable 5, GPT-5.6 Sol, and Opus 4.8 on the benchmarks it chose to publish. Full weights land July 27, which is when the claims stop being marketing and start being verifiable.

The architecture is doing a lot of the talking. K3 fires 16 of 896 experts per token, roughly 1.8% of the pool, and pairs that with what Moonshot calls Kimi Delta Attention and a 1-million-token context window. Tom's Hardware puts the parameter count at 2.8T; Fortune, citing Moonshot's own release notes, says 2.7T. The gap is small but instructive about how much of this is still self-reported.

On Arena.ai's Frontend Code evaluation, K3 landed first at 1,679 points. Moonshot's aggregate positioning has it in the top three on the benchmarks it released, with a claimed 2.5x scaling efficiency gain over Kimi K2.

The pricing is the real signal, and it cuts against the usual Chinese-model narrative. API access runs $0.30 per million cache-hit input tokens, $3 per million on cache-miss input, and $15 per million output. That uncached input price is five times what K2 charged at launch, when the tier was $0.60. Moonshot is no longer running the volume play. K3 sits well below Fable 5's $50-per-million output tier, but it's clearly aimed at the same buyer, not the hobbyist.

That buyer already exists. CNBC and Fortune both note that Cursor, DoorDash, and Thinking Machines have used prior Kimi releases in production, and Andy Fang, DoorDash's CTO, has publicly discussed the K2 lineage. Bank of America analyst Alex Liu is tracking the pricing shift as evidence that Moonshot thinks it has an enterprise product.

Hovering over all of it's February's accusation from Anthropic that Moonshot trained on 3.4 million Claude exchanges via distillation, a charge Moonshot has never conclusively rebutted. Shipping the largest open-weight model ever released, one week before dropping the weights, on the heels of that dispute, is its own form of narrative management. If the weights hold up, the distillation question becomes retrospective. If they don't, it becomes the story.

Nvidia, whose export-constrained silicon still underwrites most of this training, gets a citation in the Tom's Hardware coverage and a reminder that the frontier isn't quite as bifurcated as Washington keeps insisting. K2.5 and K2.6 were incremental. K3 is a posture change.

Sources