Kimi K3

Spec

Attribute	Value
Developer	Moonshot AI
Released	2026-07-16
Announced	2026-07-16
Context window	1M tokens
Pricing	$0.30/M cache-hit input · $3.00/M cache-miss input · $15.00/M output
License	Modified MIT (commercial use permitted; weights publicly downloadable)
Availability	Kimi Code, Kimi app, Kimi API; weights on Hugging Face (moonshot-ai/kimi-k3, released 2026-07-27); Amazon Bedrock from 2026-09-18

Release Date

July 16, 2026.

Benchmarks

• Frontend Code Arena: beats Anthropic Fable 5 (human-preference Elo) — Moonshot's reported benchmark
• Positioned as strongest open-weight model at launch for agentic coding

Use Cases

• Long-horizon agentic coding (K3 Swarm Max variant)
• Chat and general assistant tasks (K3 Max variant)
• Parallel agent workloads (K3 Swarm Max)

Availability changes after release

• 2026-09-18 — Kimi K3 becomes available on Amazon Bedrock, giving a hosted API option alongside the weights. Captured 2026-09-30, +12 days, on the routine's Moonshot rotation slot; the release itself is not new and only the surface changed. One search pass, no first-party read — huggingface.co and Moonshot's own hosts are blocked from this pipeline.
• 2026-09-08 — Unsloth published an updated K3 usage guide aimed at running the model without a multi-GPU cluster. Noted because it bears on who can actually serve a 2.8T open-weight model, and recorded as a third-party artefact.

No Kimi K4 exists. The rotation check found only a single report — The Information, around 2026-07-29 — that Moonshot is seeking additional Blackwell-class GPU capacity. No name, parameter count, benchmark, price or release date has come from Moonshot, and nothing is recorded here as forthcoming.

Compared To

Model	Params	Context	Open?
Kimi K3	~2.8T MoE	1M	Yes (Modified MIT, released July 27)
Qwen 3.8 Max	~2.4T MoE	unknown	Planned
GLM-5.2	744B MoE	1M	Yes (MIT)
Claude Fable 5	unknown	200K	No
MiniMax M3	428B MoE	1M	Yes

Architecture

• Kimi Delta Attention (KDA): hybrid linear attention mechanism enabling efficient 1M-token processing
• Attention Residuals: enhances long-context coherence
• 896 experts; ~16 active per token (~1.8% activation ratio)
• Native visual understanding

Variants

• Kimi K3 Max — chat and agent tasks
• Kimi K3 Swarm Max — large-scale parallel agent processing

Sources

• Kimi API docs: quickstart
• Simon Willison analysis: kimi-k3
• Tom's Hardware: largest open-weight