Image: Kimi

Moonshot AI's next big Kimi model is further along than you might think, but it isn't ready yet. According to leaker Lentils80, Moonshot's next flagship, internally codenamed "x4", has already finished pre-training and is now in mid-to-late post-training, with core capabilities still being trained in.

🚨 Leak:

Don't expect the next major Kimi model to be very soon

Some sources tell me Moonshot's next flagship, internally codenamed "x4", already finished pre-training. However, it's still in mid-to-late post-training, with core capabilities still being trained in

— Lentils (@Lentils80) August 12, 2026

Lentils80 shared the leak back in August, citing unnamed sources, with a blunt message: don't expect the next major Kimi model very soon. Two months on, that call still holds. Moonshot's API platform changelog shows no new flagship since Kimi K3 arrived in July, and K3 is still the model the company tells developers to move to.

To see why the leak matters, it helps to know how these models are built. Pre-training is the giant first pass, where a model reads huge amounts of text and picks up general knowledge. Post-training comes after. It's where a lab teaches the model to follow instructions, reason step by step, write code, use tools and behave safely.

Finishing pre-training is a big milestone, but post-training is where a raw model becomes something people can actually use. If core skills are still being trained in, as Lentils80 says, there's real work left before any launch. It also means it's far too early to say how good the new model will be.

Lentils80 didn't say what the model will be called. Still, the leak lines up with earlier reporting. In late July, The Information reported that Moonshot was planning a bigger successor called Kimi K4 and wanted more Nvidia Blackwell chips to train it, according to Investing.com. Whether "x4" and K4 are the same model hasn't been confirmed.

Moonshot has good reason not to rush. Kimi K3 launched as a 2.8-trillion-parameter model with a 1-million-token context window, and Moonshot calls it the first open model in the 3-trillion-parameter class. Its full weights followed about 11 days later. Even so, Moonshot's own blog admits K3 still trails Claude Fable 5 and GPT 5.6 Sol overall, with a "noticeable gap in user experience." A successor has to close that gap to make a splash.

We covered K3 when it launched, including a coding test against Fable and GPT-5.6 Sol where it scored well but took about three times longer per task.

Moonshot's release history also points to patience. Kimi K2 arrived in July 2025, and K3 came a full year later. In between, the company shipped point updates instead, with K2.5 in January, K2.6 in April and the coding-focused K2.7 Code in June. Going by that pattern, a smaller update is more likely than a whole new generation in the near term.

Kimi users aren't stuck waiting with nothing, either. K3 is available on Kimi.ai, the Kimi apps, Kimi Code and the API, and Moonshot added new web search tools and caching options for developers in September.

Lentils80 has a solid track record, though Moonshot hasn't confirmed any of this and training timelines can slip. For now, the safest bet is that Kimi's next big model is coming, just not tomorrow.