Kimi K3 Release Eve: What We Actually Know (and What Is Still Leak Noise)
Kimi K3 is not officially released as of July 16, 2026. Moonshot AI's public surfaces still treat Kimi K2.6 and Kimi K2.7 Code as the live flagships. What has arrived is a dense stack of pre-release signals: a short-lived Open Platform promo page, a high-traffic official teaser video, community arena sightings, and a recycled set of scale numbers (~2.5T parameters, ~1M context) that still lack a model card.
This post is a release-eve briefing for people searching Kimi K3, Kimi K3 release date, or Moonshot K3 leak. It separates confirmed facts from medium-confidence leaks and low-confidence hype, so you can update the page the moment Moonshot ships real docs.
Living note: If Moonshot publishes a model card, Hugging Face weights, or a public API model ID after this article is posted, treat those primary sources as authoritative over every number in the tables below.
TL;DR
| Question | Best current answer | Confidence |
|---|---|---|
| Is Kimi K3 out? | No. No public model card, weights, API ID, or official benchmark table. | High |
| What can you use today? | Kimi K2.6 (general / agentic) and Kimi K2.7 Code (coding-focused). | High |
| Why is everyone talking about July 15? | A K3 launch top-up promotion briefly appeared on Moonshot's own Open Platform, then was removed. That is a promo window, not a shipping receipt. | High that the page existed; medium that it maps to GA day |
| Rumored scale? | Community + secondary reports: ~2.5T MoE, ~1M context, "new architecture" vs a pure K2 scale-up. | Medium — single-source style leaks, widely copied |
| Official teaser? | Yes. @Kimi_Moonshot posted a teaser video on July 15 that the community reads as a K3 countdown. | High that the video exists; medium that it locks a specific ship date |
| Arena codename? | Anonymous model "Kivine" is widely assumed to be K3. Not confirmed by Moonshot. | Medium / weak for the identity claim |
| Next calendar bet? | After July 15 passed without a model card, many trackers shifted to the July 17–20 WAIC 2026 window. | Speculative |
Why this moment matters
Moonshot spent a year building a recognizable product story:
- Open weights + aggressive pricing on the K2 line (Modified MIT-style licensing on prior flagships).
- Agentic coding and long-horizon execution — multi-hour runs, tool swarms, Anthropic-compatible APIs for coding agents.
- A near-monthly cadence in 2026 — K2.5 (multimodal), K2.6 (agent production), K2.7 Code (coding efficiency).
K3 is the first time the brand has to prove a generation jump, not an iteration. That is why the leak cluster hit so hard: markets and developers are already comparing Chinese open-weight leaders (DeepSeek V4 line, GLM 5.x, and others) against closed frontier models. A credible K3 would be a pricing and infrastructure event, not only a leaderboard event.
For site readers: we already track the waiting room on the Kimi K3 Status page. This article is the long-form companion — denser sourcing, clearer uncertainty labels, and a post-launch checklist.
Timeline: from funding to teaser
| When | What happened | How to read it |
|---|---|---|
| Jan 2026 | Moonshot closes a reported $500M Series C at ~$4.3B valuation; reporting ties capital to compute expansion and K3 development. | Strong strategic signal that K3 is a funded, named generation — not a rumor inventing a product line. |
| Apr 2026 | Kimi K2.6 GA: long-horizon coding, agent swarms, production positioning. Earlier analysis on this site treated K2.6 as infrastructure for a larger supervisor model (K2.6 as runway for K3). | Confirmed product history. |
| Jun 12, 2026 | Kimi K2.7 Code ships as the coding-specialist drop (K2.7 Code release guide). | Confirms the "ship monthly, specialize when needed" pattern. |
| Early Jul 2026 | Chinese tech aggregators report "K3 this month," citing ~2.5T params, 1M context, multimodal ambitions, and comparisons to DeepSeek / ERNIE scale. | Secondary reporting; often one insider brief, many reprints. |
| Jul 14–15, 2026 | Screenshots of a Kimi K3 launch recharge campaign on the Open Platform circulate on X (roughly Jul 15 – Aug 11, bonus credits tiered ~10–30% by top-up size). The page is later gone. | Strongest platform-side artifact so far — still not a model card. |
| Jul 15, 2026 | Official Kimi account posts a cinematic teaser (million-class views within hours). Community frames it as K3 marketing. | Strong marketing signal; no specs attached. |
| Jul 15–16, 2026 | Community posts claim beta selector access and an Arena stealth model (Kivine), plus viral frontend one-shot demos (polished product pages, single-file HTML games). | Experiential, hard to audit; treat demos as demos. |
| Jul 16, 2026 (today) | Still no public API listing for K3, no Hugging Face model card under Moonshot for K3, no official benchmark table. | The honest status line. |
The important sequencing detail: promo start ≠ GA. Moonshot has previously used top-up bonuses around model launches; a deleted or early-published promo page can still be real while the weights remain internal.
Spec sheet: confidence, not vibes
Every "Kimi K3 specs" table you see online is only as good as its sourcing. Here is a tighter version.
| Claim | Status | Confidence | Notes |
|---|---|---|---|
| Exists as Moonshot's next major generation | Confirmed as intent | High | Named in funding coverage; CEO-level K3 roadmap language in secondary reports; teaser cycle is live. |
| Public GA on July 15, 2026 | Not confirmed | Low–medium | Promo window pointed there; day passed without model card. |
| ~2.5T total parameters (MoE) | Unconfirmed | Medium | Recurring figure in Chinese briefings and Western aggregator rewrites; no primary Moonshot datasheet. |
| ~1M token context | Unconfirmed | Medium | Consistent with the K2.x pattern of pushing long-horizon work; still a rumor until API docs say so. |
| "Brand-new architecture" (not just bigger K2) | Unconfirmed | Medium | Community language around residual attention / hybrid linear variants / sparsest MoE ratios — speculative. |
| Active parameter count | Unknown | — | Critical for serving cost; still missing. |
| Multimodal scope (image / video / audio) | Rumored | Low–medium | K2.5/K2.7 already multimodal; K3 scope not documented. |
| Open weights + Modified MIT (or similar) | Pattern-based expectation | Medium | Reasonable prior from K2 family; not a promise for K3. |
| Beats Claude / GPT on coding overall | Unverified | Low | Early tester claims exist; no independent harness agreement. |
| Agent Swarm / long-horizon focus | Directionally consistent | Medium–high as product story | Aligns with K2.6 infrastructure investments; still needs K3-specific docs. |
How not to misread "2.5T"
Parameter count is a capacity and cost signal, not a quality certificate. For MoE models, the number that usually decides latency and dollar cost is active parameters per token, not total experts stored on disk. Until Moonshot publishes total and active counts (and the expert routing recipe), "largest domestic model by parameters" headlines are incomplete.
Likewise, a 1M context claim only becomes useful when you know:
- which products get the full window (chat vs API vs coding CLI);
- the price tier for long prompts;
- whether quality holds at 500K–1M or only "accepts" those lengths.
How K3 is being positioned vs the K2 family
| Model | Approx. public shape | Context (public) | Role |
|---|---|---|---|
| Kimi K2 (2025) | ~1T MoE / ~32B active | 128K → later expansions | Open-weight breakthrough generation |
| Kimi K2.5 | Multimodal agent line | Long context (K2-class) | Vision + agent swarm v1 |
| Kimi K2.6 | 1T-class MoE, agent production | ~256K–262K class | Long-horizon coding, large swarms |
| Kimi K2.7 Code | Coding specialist on K2 spine | 256K | Token-efficient coding / always-on thinking |
| Kimi K3 (rumored) | ~2.5T MoE, new arch | ~1M | Next generation supervisor / flagship |
The cleanest non-leak argument for what K3 should be was already on this site in April: K2.6 looks overbuilt for its own base model — 12-hour envelopes, large agent topologies, and context compression look like scaffolding for a stronger supervisor. K3 is the model that scaffolding was waiting for. Whether the rumored 2.5T / 1M numbers are exact is secondary to whether Moonshot ships supervisor quality that can fill the harness.
What the community is testing right now
Three clusters dominate X and Reddit discussion. None of them replace a model card.
1. The Open Platform promo leak
Multiple accounts posted screenshots of a limited-time K3 launch recharge event: date window starting July 15 (China time), bonus voucher tiers by top-up size, running into mid-August. The page disappeared. That combination — real-looking first-party UI + rapid takedown — is why "launch tomorrow" posts exploded. It is still only a commercial artifact.
2. The official teaser
Moonshot's own account posted a polished video teaser on July 15. Engagement was massive by lab standards. Teasers prove marketing readiness more than technical readiness, but they raise the cost of a long silent delay: the brand is already in countdown mode.
3. "Kivine" and frontend one-shots
Arena watchers flag a stealth model named Kivine as probable K3. Parallel posts show striking frontend generation demos: interactive product pages, single-file HTML experiments, visual polish closer to top closed coding models' "taste" demos than to dry algorithm problems.
Treat these carefully:
- Stealth arena names are not official codenames until Moonshot says so.
- One-shot demos are easy to cherry-pick.
- Several posts already note a tradeoff: strong results, sometimes slow / overthinking — exactly the kind of issue a coding launch often ships a "HighSpeed" or efficiency mode to fix (as Moonshot did around K2.7 Code).
Competitive frame (without fake scoreboards)
A useful K3 comparison set, once numbers exist:
- Chinese open / open-weight peers: DeepSeek V4 line, GLM 5.x family, other domestic flagships racing scale and coding agents.
- Closed coding frontiers: Claude Opus-class and GPT-class coding stacks that still set the bar on hard repo tasks and tool reliability.
- Moonshot's own last mile: K2.7 Code is the honest baseline. If K3 cannot clearly beat K2.7 Code on long agent sessions and improve economics, the generation label is marketing.
What we will not do in this article: paste a fake "K3 vs Claude vs GPT" leaderboard from screenshots without harness names, seeds, and dates. That content ages into misinformation within hours of a real release.
What you should use today
Until K3 is public:
- General agentic work / multimodal workflows: start from Kimi K2.6 product surfaces and API listings.
- Software engineering agents and Kimi Code: use Kimi K2.7 Code (
kimi-k2.7-codeon the official API docs) — see our Kimi Code + K2.7 guide. - Migration hygiene: keep prompts and tool adapters on Anthropic-compatible shapes where possible so a K3 drop is a model-ID swap, not a rewrite.
- Status page: bookmark Kimi K3 Status for a shorter "is it out yet?" loop.
Self-hosting a multi-trillion-parameter MoE, if K3 open-weights land, will be a cluster problem, not a single Mac Studio problem — even if quantized. Plan capacity only after the weight format and active-parameter counts are public.
Launch checklist: update this article when these land
When Moonshot ships, verify in this order:
- Primary announcement — official blog, X account, or docs page with a frozen model name.
- Model card — total params, active params, architecture notes, context limits, modalities, license.
- Access paths — consumer app, Open Platform model ID, pricing per million tokens, any open weights (Hugging Face org + license text).
- Independent evals — SWE-bench / SWE-Bench Pro style coding, Terminal-Bench or equivalent, agent tool-use harnesses, and at least one third-party arena with a named identity (not only stealth).
- Efficiency story — thinking-token cost, speed tiers, cache pricing. Capability without economics does not ship into production fleets.
- Delta vs K2.7 Code — same prompts, same repos, same tool budget. Generation jumps should be visible without marketing adjectives.
If you only track three URLs after launch, make them: Moonshot/Kimi official docs, the weight host (if any), and one independent eval leaderboard you trust.
FAQ
Is Kimi K3 released?
No, not as a public, documented product as of July 16, 2026.
What is the Kimi K3 release date?
There is no official date. The strongest artifact was a platform promo window starting July 15, 2026; that day did not produce a public model card. Community attention has partly shifted toward the WAIC 2026 (July 17–20) calendar as a plausible stage, but that remains a bet.
How many parameters does Kimi K3 have?
Unconfirmed. Leaks and secondary reports often cite ~2.5 trillion total parameters in a MoE design. Treat every "2.5T" headline as provisional until Moonshot publishes a model card.
Does Kimi K3 have a 1M context window?
Unconfirmed. The 1M figure is widely repeated and would fit Moonshot's long-horizon product direction, but it is not in an official K3 datasheet yet.
Is Kimi K3 open source?
Not announced. Prior K2-family flagships were released with open weights under a Modified MIT-style license. That history creates an expectation, not a guarantee, for K3.
What is Kivine?
A stealth / anonymous arena model name that community accounts associate with K3. Moonshot has not confirmed the mapping in primary docs.
Should I wait for K3 or build on K2.7 Code now?
Build on K2.7 Code / K2.6 now if you have real work. Waiting for an unreleased model is how roadmaps slip. Design for model-ID portability so K3 is an upgrade, not a rewrite.
Bottom line
The honest summary for Kimi K3 on release eve:
- Moonshot is clearly warming a next generation — funding language, teaser creative, and platform promo residue all point the same direction.
- The viral spec sheet is still a leak sheet — 2.5T, 1M context, and "new architecture" are plausible and consistent, not verified.
- July 15 was a signal, not a delivery — no model card means no ship.
- The productive move for builders is to master the K2.6 / K2.7 Code stack and keep workflows portable.
When the real card drops, the only numbers that matter are the ones Moonshot freezes in documentation — and the independent evals that survive contact with other people's repos.
Disclaimer: kimi-k2.org is an independent informational site, not the official Moonshot AI or Kimi product site. This article synthesizes public web reporting, official Moonshot/Kimi product history, and public social posts as of July 16, 2026. Pre-release specifications can change or prove false. Always prefer primary Moonshot documentation over secondary summaries — including this one — after launch.