Kimi K32.8T flagship — specs, price, API, when to switch
Kimi K3 is Moonshot AI's flagship model, released July 16, 2026: about 2.8 trillion parameters, a 1-million-token context window, native multimodal input, and always-on thinking. It is available on Kimi.com, Kimi Work, Kimi Code, and the Kimi API (model id kimi-k3).
Use K3 for hard multi-domain jobs that keep hitting 256K walls. Keep K2.7 Code for day-to-day IDE shipping, and K2.6 when you want a cheaper, proven long agent. This site is independent of Moonshot — re-check official docs before you spend real money.
Looking for release-date timelines and source notes? Open the Kimi K3 status page
Meet Kimi K3 (official video)
Watch Moonshot's launch overview, then scroll for independent specs, pricing, and switch advice on this page.
Embedded official launch video from the Kimi AI YouTube channel. Opens externally on YouTube
Related paths
From Kimi K3 to status, Code, and K2.7
Use the flagship hub, then jump to release status, coding workflows, and older model context.
Kimi K3 Status
K3 availability tracker with official Kimi Code model docs and plan notes.
ExploreKimi Code
Setup paths, API notes, and developer workflow guidance for Kimi Code.
ExploreKimi K2.7 Code
Current release overview, capabilities, and practical context for K2.7 Code.
ExploreKimi K2.6
K2.6 overview for readers comparing older release context with K2.7.
ExploreKimi K3 pricing
Public API list prices around launch. Cache hits can dominate the bill on coding workloads.
Per 1M tokens — strong on repeated context
Per 1M tokens
Per 1M tokens — watch long reasoning traces
Prices change. Confirm on platform.kimi.ai before production spend. Membership plans for Kimi Code may use different entitlements than raw API pricing.
Kimi K3 specs & architecture
The numbers people search for, plus the architecture names you will see in launch posts (KDA, Attention Residuals, Stable LatentMoE).
Kimi K3 features & capabilities
Flagship general intelligence — not a free upgrade for every autocomplete click.
Sustains long engineering sessions, navigates large repos, and iterates with tools, logs, and runtime feedback.
Strong when the job spans many files and hours, not only a single patch.
Thinking stays on at launch (max effort). Built for hard multi-step plans rather than shallow chat replies.
Prefer a new session when switching models so thinking history stays clean.
Images and video in the same model loop — useful for UI screenshots, game frames, and visual debugging.
API inputs typically use base64 or platform file IDs, not arbitrary public HTTP image URLs.
Research packs, slides, dashboards, and long agent runs that mix search, code, and synthesis.
Best as a supervisor / hard-problem model, not the default for every small task.
Kimi K3 benchmarks (summary)
Official launch charts for coding and agent suites. Treat them as vendor-reported; re-check independent leaderboards over time.
Moonshot positions K3 near frontier proprietary models on several internal and public coding/agent evals, while still trailing the top closed models overall on some suites. For a narrative read of what that means for switching traffic, open the release deep-dive.
How to use Kimi K3
Four practical entry points — app, Code, API, and future open weights.
Update the Kimi app or open kimi.com and pick K3 in the model picker when available.
Best for chat, agents, and everyday flagship use.
In Kimi Code, select K3 via the model command. Product docs may list the ID as k3 (plan limits apply).
Membership tier can still block access or 1M context.
Wire model id kimi-k3 on the official Kimi / Moonshot API platform.
List price band above; verify live pricing before production.
Moonshot said full model weights would follow after launch (around July 27, 2026).
Until HF/card is live, “open” is a roadmap word — not a folder on your disk.
Kimi product surfaces do not always share one string. Use the ID that matches the product you are calling.
Use kimi-k3 for OpenAI-compatible API calls and pricing pages.
Official Kimi Code model docs list K3 as k3, with plan-gated context (up to 1M on higher tiers).
Kimi K3 vs K2.7 Code vs K2.6
Newest is not automatically best for every ticket. Pick by pain, not by logo.
Flagship general intelligence: 1M context class, premium price, hard multi-domain + long sessions.
Switch when you keep hitting context walls or need max-effort reasoning on tough work.
Coding specialist for IDE/CLI loops: Kimi Code, Claude Code-compatible flows, Cline, RooCode. ~256K context.
Still the boring correct default for shipping PRs day to day.
Proven general long agent at a lower list-price band than K3.
Keep it for chores that already work well without a flagship bill.
Quick decision tree
- →Hitting 256K walls or need flagship multi-domain depth → pilot kimi-k3.
- →Daily coding agents / PR loops → stay on K2.7 Code first.
- →Long autonomous chores already stable on K2.6 → do not upgrade only for the logo.
- →Need local open weights today → wait for the announced dump; use API/product in the meantime.
Kimi K3 limitations (read before you bet a sprint)
From public launch notes and practical agent harness experience — not reasons to ignore K3, reasons to configure it.
K3 was trained with preserved thinking history. Harnesses that drop intermediate thinking, or mid-session model switches, can make quality unstable.
Prefer a verified harness (e.g. Kimi Code) and start a new session when switching models.
Long-horizon training can make K3 improvise when intent is ambiguous — sometimes past the boundary you wanted.
Tighten system prompts / AGENTS.md when you need strict boundaries.
Public materials still place overall UX behind the top proprietary models on some dimensions.
Pilot hard jobs; do not flip all production traffic on day one.
Kimi K3 FAQ
Release status, pricing, API id, open weights, OpenRouter, and when to stay on K2.7 Code.
What is Kimi K3?
Kimi K3 is Moonshot AI's flagship model released July 16, 2026. Public materials describe about 2.8T parameters, a 1M-token context window, native multimodal input, and always-on thinking for long-horizon coding and knowledge work.
Is Kimi K3 released?
Yes. As of July 16, 2026, K3 is available on Kimi product surfaces and listed for the Kimi API. For a release-date timeline and source notes, see the Kimi K3 status page on this site.
What is the Kimi K3 release date?
Public product/API availability is documented from mid-July 2026 (July 16 launch materials). Older May/June 2026 rumor dates were not confirmed as public model-doc listings.
What is the Kimi K3 API model ID?
On the Kimi API platform the model id is kimi-k3. In Kimi Code documentation the selectable model may appear as k3. Always use the ID from the product you are integrating.
What is Kimi K3 pricing?
Public launch API pricing has been listed near $0.30 per 1M tokens for cache hits, $3 for cache-miss input, and $15 for output. Prices change — confirm on platform.kimi.ai before production spend.
Should I switch everything from K2.7 Code to K3?
Usually no. Keep K2.7 Code for day-to-day coding agent loops. Pilot K3 on hard, long, multi-domain jobs or when 256K is not enough. K3 is a flagship tier, not a free rename of K2.7.
How large is the Kimi K3 context window?
Official materials list 1,048,576 tokens (1M). Access to the full window may still depend on product surface and membership plan.
When will Kimi K3 open weights ship?
Moonshot's launch materials said full model weights would be released around July 27, 2026, with a technical report to follow. Treat that date as an official plan until the weights page is live.
Can I use Kimi K3 on OpenRouter or third-party routers?
Third-party routers often list Moonshot models shortly after launch (search for moonshotai/kimi-k3 or similar). Always verify the live model slug, price, and rate limits on that router — do not assume they match the official API card.
How does Kimi K3 compare to Claude or GPT flagship models?
Moonshot's launch materials position K3 near frontier proprietary models on several coding and agent evals, while still trailing the very top closed models overall on some suites. Independent leaderboards will move; use vendor charts as a starting point, not a permanent ranking.
Is this the official Kimi website?
No. kimi-k2.org is an independent unofficial resource. For authoritative specs, pricing, and downloads, use moonshot.ai, kimi.com, and platform.kimi.ai.