Kimi K3 Context Window: 1 Million Tokens, Confirmed

What was a rumor in March is now official: Kimi K3 supports up to a 1 million token context window. But the full window isn't available to everyone — here's how it actually works.

The Tier Breakdown

Plan	K3 context window
Andante	K3 not available
Moderato	256K tokens
Allegretto and above	1M tokens

For API and third-party tool usage, set the context-window field to 1048576 to unlock the full million. Accounts that exceed their plan's limits get a 401 response.

What Is a Context Window?

A context window defines how much content a model can "see" at once — its working memory:

• 4K tokens ≈ 3,000 words — a short article
• 128K tokens ≈ 96,000 words — a full novel
• 1M tokens ≈ 750,000 words — seven full novels simultaneously

Kimi K2 topped out around 256K. K3's jump to 1M is a ~4× increase.

What Changes at 1 Million Tokens

For Developers

• Code review at scale — analyze entire repositories without chunking
• Documentation generation — pass the full source tree and get accurate, repo-aware docs
• Dependency analysis — understand inter-module relationships directly

For Enterprises

• Process full legal contract bundles in one shot
• Cross-reference a year of meeting transcripts
• Compare multiple research papers simultaneously

For End Users

• Upload and chat with an entire book
• Persistent conversations that never lose early context
• Interactive analysis of large datasets or log files

How K3 Actually Pulls It Off: KDA

The pre-launch speculation guessed at "sparse attention." The real answer is more interesting: Kimi Delta Attention (KDA), a hybrid linear-attention mechanism combined with attention residuals.

Standard transformer attention scales as O(n²) with context length — ruinous at 1M tokens. KDA replaces most of that quadratic work with linear-complexity state updates (a delta-rule mechanism in the family of linear attention / gated delta nets), while keeping enough full-attention capacity — via the hybrid layers and attention residuals — to preserve retrieval quality deep in long contexts. It's the architectural reason K3 can offer 1M tokens as a production feature rather than a benchmark stunt.

K3's MoE design (2.8T total parameters, but only 16 of 896 experts active per token) keeps per-token inference cost manageable at long contexts, and the Mooncake serving architecture caches prompt prefixes aggressively — coding workloads see 90%+ cache hit rates, which is also what keeps long-context sessions affordable.

Kimi's Long-Context Heritage

This isn't Moonshot AI's first long-context rodeo. Kimi was one of the first commercial AI products to offer 200K context back in 2024, while most competitors were still capped at 32K. K3 doubles down on that identity — with the architecture to back it up this time.

Use It Now

• kimi.ai — consumer product (Moderato+ for K3)
• platform.kimi.com — API for developers

For cost implications of long contexts, see our Kimi K3 pricing breakdown.