Kimi K3 soft-launches with up to 1M context — official specs still pending

On the evening of July 16, 2026, Kimi’s next flagship — K3 — began showing up as a live product option, according to reporting that cites Kimi’s own in-product copy. IT Home describes it as Moonshot’s strongest Kimi model to date, positioned for coding, games / 3D, and knowledge workloads, with a headline context window of up to 1M tokens.

What has not arrived yet is equally important: as of that report, Moonshot’s Weibo, website, and WeChat channels still lacked a formal press release, and parameter counts, API pricing, and public benchmarks remain unpublished.

Key takeaways

  • Soft launch, not a full datasheet. K3 is selectable in product; the paperwork (model card, price list, open-weight decision) is still missing.
  • Sits beside Kimi K2.7 Code. Kimi Code currently lists K3 and K2.7 Code — three model IDs total — so teams can keep the June coding build while trying the new flagship.
  • 1M context is membership-gated. Per the same product copy: Andante does not support K3 yet; Moderato gets 256K; Allegretto and above can reach 1M.
  • Community names in circulation: K3 Max (“chat & agent”) and K3 Swarm Max (batch / massive search). Treat variant naming as product UI until Moonshot publishes a canonical API ID list.
  • Ignore rumor numbers for now. (Confirmed July 27: 2.8T-parameter MoE with 104B active. See the Kimi K3 model page.)

What appears to be live

SignalStatus (as of July 16 evening reports)
In-product K3 option (web / client / Code tooling)Reported live
Up to 1M context on higher membership tiersReported in official product copy
Formal Moonshot press release / blogNot yet
Official model card / benchmark tableNot yet
Confirmed API pricingNot yet
Open weights on Hugging FaceNot yet (prior K2-series precedent ≠ a K3 promise)

Earlier on July 15–16, Moonshot posted teaser videos and the community circulated head-to-head clips against Claude Fable 5 and GPT-5.6 Sol. Those are marketing/demo signals — useful for intent, not for scoreboards.

Why it matters

  1. China’s open-and-hosted coding race just got another flagship signal. Even without a card, shipping K3 next to K2.7 Code tells developers Moonshot is moving the default target beyond the K2.x line.
  2. Long context remains a Kimi brand bet. Tiering 256K vs 1M by membership mirrors how other vendors monetize extreme windows — check your plan before assuming every seat gets 1M.
  3. EEAT discipline: until Moonshot posts numbers you can cite, do not plan capacity, price, or “beats Opus” claims on leaked parameter counts.

Update: K3 is now fully launched

July 27, 2026: Moonshot AI released the full Kimi K3 model weights (2.8T parameters, 104B active MoE), a 47-page technical report, and infrastructure code on GitHub and Hugging Face. API pricing is $3/$15 per 1M tokens. The model ships under a custom Kimi K3 License (not MIT/Apache). See the full Kimi K3 model page for complete specs, benchmarks and the license details.

For landscape context, see the 2026 LLM guide.

Sources

← All news