Kimi K3 Token Hub

What is Kimi K3?

Last updated:

Kimi K3 is a 2.8-trillion-parameter open-weight AI model from Moonshot AI, announced on July 16, 2026. It accepts text and images, handles a 1,048,576-token (1M) context window, and costs $3 per million input tokens / $15 per million output tokens via the official API. Moonshot has promised to publish the weights by July 27, 2026, which would make K3 the largest open-weight model ever released.

Kimi K3 specs overview 2026: 2.8T parameters, 1M token context window, $3/$15 per million tokens API pricing, open weights due July 27

Key specifications

AnnouncedJuly 16, 2026 (API listing dated July 15)
Parameters≈2.8T total, mixture-of-experts; active count undisclosed
Context window1,048,576 tokens (1M)
ModalitiesText + image input, text output
Architecture notesKDA (Kimi Delta Attention) plus attention residuals, per the OpenRouter listing
API price$3 / 1M input tokens · $15 / 1M output tokens
LicenseOpen weight — weights promised by July 27, 2026

Why Kimi K3 matters

K3 is Moonshot's attempt to put an open-weight model in the same conversation as the strongest closed frontier models. Independent evaluator Artificial Analysis reported an Elo of 1547 on its private evals — a jump of roughly 732 points over Kimi K2.6 — with K3 mostly outperforming Claude Opus 4.8 (max) and GPT-5.5 (high) on Moonshot's self-reported benchmark table, while still trailing the newest closed flagships. Mainstream coverage from Bloomberg, CNBC and Forbes framed the release as China's open-model ecosystem closing the gap with US labs.

Two practical details stand out for developers. First, despite being much larger than K2.6, K3 is reported to use about 21% fewer output tokens to complete the same tasks — which partially offsets its higher per-token price, since output tokens are the expensive side ($15/M). Second, the 1M context window plus image input makes it a genuine candidate for repository-scale coding agents and document-heavy research workflows, the two workloads our cost calculator models by default.

How Kimi K3 compares on price

ModelInput / 1M tokensOutput / 1M tokensContextLicense
Kimi K3$3.00$15.001,048.576Kopen-weight
Kimi K2.6$0.95$4.00262.144Kopen-weight
DeepSeek V4 Pro$1.10*$4.40*262.144Kopen-weight
GPT-5.5$2.50*$20.00*400Kproprietary
Claude Sonnet$3.00*$15.00*1,000Kproprietary

* listed price at time of verification; check the linked source for the vendor's current rate.

At $3/$15, K3 is priced about 3× above its predecessor K2.6 ($0.95/$4) and lands in the same tier as Anthropic's Claude Sonnet. Whether that premium is worth it for your workload depends almost entirely on token volumes and cache behavior — see the full pricing breakdown or run your own numbers in the calculator.

How to access Kimi K3 today

  • Chat: kimi.com and the Kimi mobile apps (free tier available).
  • API: the Moonshot/Kimi platform (OpenAI-compatible endpoints) — see our API guide.
  • Routers: OpenRouter lists K3 with pass-through pricing to Moonshot.
  • Self-hosting: not yet — weights are due by July 27, 2026. Our local deployment guide tracks the release and what hardware you'd realistically need.
Kimi K3 release timeline: announced July 16 2026, open weights promised by July 27, community quantizations expected in the following weeks

What's next

The open-weights drop is the next milestone: it will trigger third-party quantizations (GGUF, MLX), independent benchmark reproduction, and cheaper hosted offerings from GPU clouds. We track those events on the blog and keep the numbers on this page current — the "last updated" stamp above tells you when a human last verified them.

Frequently asked questions

When was Kimi K3 released?

Moonshot AI announced Kimi K3 on July 16, 2026. It became available immediately on kimi.com and through the Moonshot/Kimi API; open weights were promised within roughly ten days of launch (by July 27, 2026).

How many parameters does Kimi K3 have?

Kimi K3 has about 2.8 trillion total parameters in a mixture-of-experts (MoE) design — Moonshot markets it as the first open model in the 3T class. The number of active parameters per token has not been officially published.

Is Kimi K3 multimodal?

Yes. Kimi K3 accepts image input alongside text, a first for Moonshot's K-series flagship line, and is reported to be strong at reading screenshots, logs and UI mockups.

Who makes Kimi K3?

Kimi K3 is built by Moonshot AI, a Beijing-based AI lab founded in 2023, known for the Kimi chatbot and the open-weight Kimi K2 model family that preceded K3.

Sources