Kimi K3

One-line positioning

Moonshot's open-source flagship — 2.8T params, the world's largest open model, 1M context, a domestic tech-flow benchmark

Kimi K3 is Moonshot AI's flagship model released July 17, 2026. It has 2.8 trillion total parameters (MoE, 896 experts with 16 active per inference) and a 1M-token context window, the largest open-source model by parameter count globally. It natively supports vision, uses in-house KDA hybrid linear attention and AttnRes, with training efficiency up ~2.5× over the previous generation. On the Code Arena frontend benchmark it hit 1679, beating Claude Fable 5's 1631 and GPT-5.6 Sol's 1618.

Getting Started (based on community-tested notes, not first-person)

Note: compiled from official docs and community testing, rewritten and condensed by editors — not a claim that we tested each point ourselves.

1. Use Kimi directly: web, app, and Kimi Work all chat, but from 2026-07-19 new C-end subscriptions are paused for compute shortage; existing users are unaffected.
2. Developers use API: input $3/M tokens (~¥20, cache miss), output $15/M (~¥100); cache-hit input is only ¥2. On the Mooncake architecture, programming cache-hit rate exceeds 90%.
3. Self-host if wanted: full weights, tech report, and Infra (MoonEP/FlashKDA/AgentEnv) open-sourced 2026-07-27, self-hostable.

Real Pain (community consensus)

Note: recurring pain points reported across the community, condensed by editors, not individually verified by us.

• Priciest open source domestically: output ¥100/M tokens, nearly 4× K2.6's $4, hard for budget-sensitive teams.
• C-end subscriptions paused: from 2026-07-19 new-user subs are paused, all membership tiers "sold out"; you wait for compute expansion.
• Name confusion: K3 is the model; Kimi is the product — they get mixed up, keep them distinct in proposals.
• Hallucination up: independent tests cite AA-Omniscience hallucination at 51%, up from K2.6's 39%, not improved [verify].
• High deploy bar: full 2.8T inference needs large VRAM clusters; individuals mostly use API or quantized versions.

Hard Comparison (one-line verdicts)

• DeepSeek: another domestic open benchmark, cheaper and louder community; Kimi wins on 1M context and frontend top spot, but costs more per token.
• Doubao: ByteDance's 300M MAU entry and cheap API; Kimi wins on tech depth and open weights, fit for teams wanting self-host.
• Claude / GPT: closed flagships edge the overall ceiling; Kimi wins on self-host and price-parity signaling, gap down to 3.3%.

Who It's For

• For: 1M-context research/legal/code work, enterprises wanting private deployment, tech-flow developers.
• Not for: ultra budget-sensitive API-only users (pick DeepSeek), light C-end chatters (subs paused), or those avoiding heavy models.

Editor's Take

Note: our view from an open-source ecosystem and engineering angle.

Kimi K3's symbolic weight exceeds its parameters — it's the first time domestic open source stood on the same global pricing stage as Claude and GPT; the high output price is itself a signal of "daring to price." 1M context is a real need for long docs and codebases, and topping frontend coding proves it's not benchmark padding. But the cold water: paused C-end subs show compute is genuinely tight, you may not get access now; ¥100/M output is a real cost at high volume. Our advice: researchers/enterprises validate with open weights privately first, developers use API but always enable cache-hit (programming hit rate over 90%, unit cost cut to 1/10); don't let "world's largest open model" dizzy you — calculate your call volume and bill before picking a main model.

Ratings

Parameter scale and open value: ★★★★★

Long-context ability: ★★★★★

Coding / tech-flow performance: ★★★★★

Value for money (output pricey): ★★★☆☆