Sponsored by
AI MODELS · CHINA TECH ·
Hey readers,
A launch page went live by accident this week. It sat on Moonshot AI's developer platform for a few hours before someone pulled it down — but not before half of X had already screenshotted the parameter counts nobody was supposed to see yet.
That's how most of the world found out about Kimi K3.
THE QUICK HIT
A promo page leaked Moonshot's next flagship a day before its own launch
Leaked specs point to a ~2.5T parameter model with a 1-million-token context window
Nothing is confirmed yet — K3 is still an unofficial beta, not a shipped product
The Leak Nobody Planned
On July 14, a promotional page on Moonshot's Kimi Open Platform briefly went live, advertising a "K3 launch" recharge bonus tied to a midnight Beijing-time release the next day. It was pulled within hours. It didn't matter — the internet doesn't forget a leak that fast.
Two days later, the official Kimi account posted a short teaser video built around the number three, quietly confirming what everyone already suspected: K3 is real, and it's close.
Here's what's floating around based on early leaks and beta access — none of it officially confirmed by Moonshot yet:
Architecture: Mixture-of-Experts, roughly 2.5 trillion total parameters — more than double K2.6's 1 trillion
Context window: A native 1-million tokens, built on Moonshot's own "Kimi-Linear" attention tech, claimed to cut memory overhead by 75%
Early testing: Arena testers (codename "Kivine") rate it favorably against GPT-5.6 Sol and Anthropic's Fable 5 on coding and multimodal tasks — though some report painfully slow runs on harder agent jobs
No model card. No official benchmarks. Moonshot hasn't confirmed a single spec. Treat every number above as leaked, not shipped.
Why Moonshot Needed This Win
This isn't just a routine model refresh. It's a comeback attempt.
Kimi K2.6 launched in April and briefly became the most downloaded open-weight model anywhere — until Z.ai's GLM-5.2 took the crown in late June. Losing that spot stung for a company that built its identity around long-context, open-access AI.
$500M — Moonshot's Series C, closed January 2026 at a $4.3B valuation, reportedly earmarked specifically for K3's training run.
That's not a company hedging its bets. That's a company betting the roadmap on one model landing hard.
Quick Word From Today's Sponsor
What do Ridge, Salt & Stone, and Madhappy have in common?
They use Shotpin to capture marketing consent. Shoptin replaces the default with compliant, region-optimized consent UX: GDPR-correct in the EU, TCPA-safe in the US, persuasive everywhere. New contacts sync straight into Klaviyo, Attentive, or Postscript. Zero code setup. 30-day completely free A/B test.
Continue Reading
Open Weights, Still Unconfirmed
Here's where it gets interesting — and where a lot of coverage is getting ahead of itself.
Moonshot released K2 under a Modified MIT license, making the weights freely downloadable and fine-tunable. Most reporting assumes K3 will follow the same pattern. But Moonshot hasn't confirmed that yet. Until the weights actually drop, "open" is an expectation, not a fact.
If it does ship open, the implications reach past one model. Open weights let companies inspect, customize, and self-host instead of renting intelligence by the token — a fundamentally different economic relationship than the one closed labs offer.
What It Means For Builders
Whether or not K3 ends up open, the comparison already tells a story:
Model | Status | Parameters |
|---|---|---|
Kimi K2.6 | Live since April 2026 | ~1T total, 32B active |
GLM-5.2 | Open-weight leader | 753B total |
Kimi K3 | Leaked, imminent | ~2.5T (unconfirmed) |
Claude Opus 4.8 | Live, closed | ~1.5–2T estimated |
Notice what's missing from that table: certainty. Every open-weight lab in this race is now leaking specs before shipping them, because pre-announcing freezes competitor attention even when the product isn't ready. That's a marketing tactic as much as a technical one.
There's a quieter thread worth watching too. Anthropic's own safety researchers have already tested Kimi K2.6 in simulated fraud scenarios as part of broader agentic-misalignment research. Bigger, faster open models with agent capabilities aren't just a pricing story — they're a growing scrutiny story, especially once businesses start handing them real tool access.
The Bigger Picture
The AI race is quietly splitting into two products: the smartest model, and the model you can actually own, inspect, and run on your own terms. Kimi K3 is a bet that the second one wins more customers over time.
Which one would you actually build on: the smarter model, or the one you control? Hit reply — I read every response.
Until the next leak,