Kimi K3: Moonshot AI unleashes an open-weight giant to sidestep the compute crunch

Chinese startup Moonshot AI has played its ace: Kimi K3. With 2.8 trillion total parameters, it's the largest open-weight model ever released — but the number itself matters less than how it's put to work. The Mixture-of-Experts architecture is sparse: out of 896 experts, only around 16 are said to fire per token, according to early analyses (Moonshot hasn't officially confirmed the figure yet). That means the computational scaffolding is enormous, but the practical footprint is surgical — a necessity dictated more by the hardware limits the sanctions impose than by pure competitive spirit.

On paper, the performance is convincing: Kimi K3 takes first place in the Frontend Code Arena benchmark, edging out models of the caliber of Claude Fable 5 and GPT-5.6 Sol. The system is multimodal — it accepts text, images and video as input — and offers a one-million-token context window. On those long contexts Moonshot claims its new Kimi Delta Attention makes decoding up to 6.3 times faster, paired with a technique called Attention Residuals. Reasoning is always on and tunable via the 'reasoning_effort' parameter, but here's where the uncertainty creeps in: the internal benchmarks may be flattering, yet independent tests on actual token consumption are still pending. The whispers about heavy resource use on trivial tasks remain, for now, just whispers.

The real strategic move, though, isn't technical but commercial. As the United States tightens the noose on advanced chip exports, China's answer isn't to compete on secrecy but on accessibility. Offering a model of this class at rock-bottom prices — $3 per million input tokens and $15 per million output — with open weights is an efficient way to route around the hardware barriers and grab market share. A shadow lingers over the license: billed as a 'modified MIT' but not yet published in full. Until July 27, when the complete weights drop, the openness stays a promise, not an accomplished fact.

Giving away access to dominate the infrastructure looks like the new capitalism. We'll see whether the promised efficiency holds up under the weight of millions of developers ready to hit download.

— Olya