Moonshot’s Kimi K3 Lands With a 2.5-Trillion-Parameter Pitch to Rival Opus 4.8

Moonshot AI released Kimi K3 on July 16, and the pitch is direct: a model built to close the distance with Anthropic’s Opus 4.8, and possibly cross it. TechCrunch reported that the Financial Times, citing anonymous sources, described K3 as the largest open-weight model to come out of China, with a parameter count somewhere between 2 trillion and 3 trillion.

That scale claim matters less on its own than what it represents: the ninth time in roughly a year that a Kimi release has reset the bar for open-source model size, starting from the original Kimi K2’s 1 trillion parameters through the K2.5, K2.6, and K2.7 Code updates that followed. Where those releases nibbled at the gap between open and closed models one point at a time, K3 looks like Moonshot deciding to jump instead of inch.

What Moonshot is claiming

Beyond scale, Moonshot says K3 delivers frontier-level performance, and that among the models it tested internally, K3’s overall intelligence trails only Claude Fable 5 and GPT-5.6 Sol. Early testers cited across several outlets place its output somewhere between GPT-5.6 and Fable 5 in practice, and describe K3 beating Opus 4.8 and GPT-5.5 on select coding benchmarks. The model reportedly uses a Mixture-of-Experts architecture and Moonshot’s own long-context technique to support a 1-million-token window, more than triple the 256K context of the prior K2.6 release.

None of that is independently verified. As of publication, Moonshot has not published an official model card, a benchmark table, licensing terms, or per-token pricing for K3. The 2.5-trillion-parameter figure that recurs across most coverage traces back to an April 2026 report out of China and has been repeated in leak after leak since, without a primary Moonshot source confirming it directly.

Kimi K2’s track record set up this moment

K3 doesn’t arrive from nowhere. Moonshot’s Kimi K2 family has performed well in the open-source market since its mid-2025 debut, ranking high on public benchmarks and staying closer to frontier closed models than most open-weight competitors managed. That track record is precisely why K3’s claims are being taken seriously rather than dismissed as marketing: Moonshot has iterated fast, shipping K2, K2.5, K2.6, and K2.7 Code inside about a year, each one closing a bit more of the distance to paid, closed-source labs like OpenAI and Anthropic.

The money behind the model

Moonshot is also reportedly raising a fresh funding round that would value the company at $31.5 billion — a sharp jump from the $20 billion valuation it reached in May, when it raised $2 billion. A model release timed alongside a valuation jump is a familiar sequence in this market: proof points for investors tend to arrive right before the round closes, which is one more reason to treat unverified specs as marketing collateral until Moonshot publishes primary documentation.

Why this lands during an open-vs-closed reckoning

K3’s launch also lands squarely inside a live argument about whether it’s worth paying for closed models at all. Microsoft CEO Satya Nadella recently warned enterprises that using proprietary AI systems means paying twice: once in subscription fees, and again by handing model providers the corrections and workflows that encode a company’s institutional knowledge. Palantir CEO Alex Karp has raised similar concerns about how AI is sold, and Solo.io CEO Idit Levine has told reporters that enterprise clients increasingly ask whether an open-source model, run on their own infrastructure, can do roughly 90% of what a closed frontier model does for a fraction of the cost.

Every open-weight release that narrows the performance gap gives that argument more weight. Whether or not K3 actually matches Opus 4.8 once independent benchmarks arrive, the fact that a credible challenger exists — one enterprises can inspect, fine-tune, and run on-premises — is the more consequential story for buyers deciding where to spend their AI budget next quarter.

The Timespek take

Treat Kimi K3 as a credible open-weight contender, not a confirmed win over Opus 4.8. Moonshot’s iteration speed and Kimi K2’s benchmark history make the claims plausible — this is not a lab making noise from nowhere. But “largest open model from China” and “beats Opus 4.8 on coding benchmarks” are two different claims resting on two different levels of evidence, and only one of them currently has a primary source behind it. Until Moonshot ships a model card and independent testers publish scores, the honest read is: watch closely, don’t take the parameter count as the headline.

Frequently Asked Questions