All posts

Kimi K3 launch tracker: what the leaks got right — and wrong

· Kimi· K3· Release Tracker

Update, July 17, 2026: Kimi K3 is out. Moonshot released it on July 16 — one day after its teaser, and a day ahead of the community's WAIC-window bets. This page began as a release-eve tracker separating confirmed facts from leak noise; now that the model is real, it becomes the scoreboard. The full breakdown of what actually shipped — specs, benchmarks, pricing, real cost math — is in the K3 launch deep dive.

The scoreboard: pre-launch claims vs. the release

Pre-launch claimStatusReality
"Kivine" on Arena is a K3 test buildEffectively confirmedKivine's generations match K3's launch profile; the community attribution held up
~2.5T-parameter MoEWrong, undersold2.8T — larger than the leak
1M-token context windowConfirmed1M (1,048,576), with flat pricing across the window
2–3T parameters, "largest open-weight model from China" (FT)On track2.8T fits the range; "open-weight" now carries a date — see below
Launch during WAIC, July 17–20Close, but earlyJuly 16 — Moonshot moved a day before the window
"Opus-class coding, behind Sol and Fable on Terminal-Bench" (leak)Directionally rightTerminal-Bench 2.1: 88.3 vs Sol's 88.8; Artificial Analysis ranks K3 #3 behind only Fable 5 and Sol
Open weights at launchDelivered — on the deadlineShipped July 27, 2026: 1.56TB on Hugging Face, under a new custom license (no longer Modified MIT) — full breakdown in the weights guide

Two lessons worth keeping from this exercise. The leaks were right about *shape* — context size, positioning, agent focus — and wrong about *numbers*, in both directions. And the one claim everyone reported as settled early ("it's open source") was the last to actually land: the weights arrived on July 27, exactly the committed date — 1.56TB on Hugging Face, with a new license attached that is no longer Modified MIT. The promise held; the terms changed. Details in the weights guide.

Where the weights stand: delivered, July 27

The commitment held to the day. On July 27, 2026 Moonshot published Kimi K3's full weights — 1.56TB on Hugging Face. The Modified-MIT expectation did not hold: K3 ships under a new custom license — free for internal use, product embedding and research; a commercial agreement required for MaaS businesses above $20M; prominent attribution required above 100M MAU or $20M monthly revenue. What that means in practice — including the brutal hardware math and the silent-quantization problem in third-party serving — is in the weights guide.

What matters now

The release-eve questions are answered; the ones that matter next are practical, and the deep dive covers them: what $3/$15 buys against GPT-5.6 Terra at the same output price, why the single "max" reasoning level makes real per-task cost higher than the sticker, and when K3 beats GLM-5.2 or K2.6 for your workload.

Kimi K2.6 and K2.7-code run today on Turiloop behind one OpenAI-compatible key; K3 will join the lineup when its API stabilizes, and the pricing calculator already carries its launch prices.

FAQ

Is Kimi K3 released? Yes — July 16, 2026, on the API and kimi.com. The open weights followed on July 27, exactly on the committed date: 1.56TB on Hugging Face.

What were K3's final specs vs the leaks? Reality: 2.8T-parameter MoE (leaks said 2.5T), 1M context (leaks were right), $3/$15 pricing with $0.30 cached input (leaks had no price). Architecture details await the official model card.

Was Kivine actually K3? The community attribution held: the anonymous Arena model's behavior matches the released K3's profile. Moonshot never confirmed it by name.

Is Kimi K3 open source? The weights are public as of July 27, 2026 — under a custom license rather than the K-series' Modified MIT: free for internal use and products, gated for large MaaS resellers, attribution required at very large scale. See the weights guide for the terms.

Where do I see K3's benchmarks and pricing? In the launch deep dive: vendor scores (GPQA Diamond 93.5, BrowseComp 91.2), third-party rankings (AA Index #3), the price ladder, and the honest cost math.