Moonshot AI's Kimi K3 Targets Opus 4.8 with 2-3 Trillion Parameters
Moonshot AI is preparing to release Kimi K3, a model it claims will match or surpass Anthropic's Opus 4.8. The release is expected within days as of July 16, 2026. If the parameter count holds, it will be the largest open-weight model to come out of China.
The Numbers
Kimi K3 sits somewhere between 2 trillion and 3 trillion parameters. For context, that range puts it well beyond most models currently available in any weight class. Moonshot has not specified where in that range it lands.
The predecessor, Kimi K2, already benchmarks close to frontier models. K3 is presumably the attempt to close the remaining gap with the top proprietary models, Opus 4.8 specifically named as the target.
Open Weight, Big Claims
Open-weight release means the weights will be publicly accessible. That matters more than the benchmark comparisons. If K3 actually performs near Opus 4.8 levels and runs on accessible hardware, that changes the economics for anyone currently paying API rates for top-tier inference.
"Claims to match" and "actually matches in deployment" are different things. Kimi K2's benchmark performance has been strong enough to take seriously. K3 with a substantially larger parameter count is a reasonable extrapolation, not pure marketing.
The Funding Context
Moonshot AI raised $2 billion at a $20 billion valuation in May 2026. It is now raising again at a $31.5 billion valuation. That is a 57% jump in roughly two months. The timing relative to the K3 release is not a coincidence.
A 2-3 trillion parameter model is expensive to train and expensive to serve. The capital raise suggests Moonshot is planning for inference costs at scale, not just a benchmark run.
What to Watch
The benchmark comparisons will land within days. The more useful signal will come from third-party evaluations on real tasks rather than standardized tests. Kimi K2 held up reasonably well under that scrutiny. K3 will get the same treatment quickly.
Open weights at this scale also means fine-tuning options that do not exist with Opus 4.8. That is a separate category of value not captured in direct benchmark comparison.
Source: Techcrunch