Kimi K3, the one that works nights.

Think about what actually happens to a group file overnight. Four supplier replies came in after you left. One quoted a different room category than you asked for. Two follow-ups are due to suppliers who went quiet. The costing draft needs updating with the new rates, and tomorrow’s summary should flag the one reply that changes the plan.

No single step here is hard. A trainee could do any one of them. The difficulty is that there are eighty steps, they depend on each other, and the file still has to make sense at 6am. Give this to most AI models and somewhere around step thirty they drift: they forget the room category issue, or chase a supplier who already answered. There is something unsettling about work happening at 3am with nobody watching, which is exactly why the model doing it has to be the kind that does not drift.

The stamina model

Kimi K3 comes from Moonshot, a Chinese lab, and its results cluster around one theme. On the scoreboard that tracks real working sessions, the kind with tools, failures and many steps, it holds fourth place, ahead of every Google and every other model outside Anthropic and OpenAI. On the scoreboard where models build complete working things from start to finish, it is second in the world.

Brilliance for five minutes is common now. Coherence at step eighty is rare.

The most telling result is a strange one. Models advertise how much they can hold in their head at once, and most quietly get worse as that memory fills up, like a desk disappearing under paper. In one published test, K3 was run on a long research task twice: once managing its memory carefully, once just letting everything pile up to the full million words. The scores were nearly identical. The desk got buried and the work did not suffer. For overnight work, where the file only grows, that is the property you want.

Speed, accuracy, cost

The three numbers. Accuracy: the best of any open model, and fourth overall, on the working-sessions scoreboard. Speed: slow, about 36 words a second, roughly a tenth of the fastest models, per Artificial Analysis. During the day, with a client waiting, that would be a real problem. At 3am it could not matter less. Cost: mid-range, about $3 per million tokens read and $15 per million written, tokens being the small chunks of text AI is billed in. Cheaper than the western flagships for work of comparable stubbornness.

The overnight batch is also where AI bills quietly balloon if you buy AI directly: eighty steps a night, every night, all billed by usage, at whatever rate your chosen model charges. Running the night shift on a premium flagship doubles the cost of your cheapest hours. One more practical point in the same direction: K3 is an open model, published for anyone to run, so many providers compete to serve it and no one can hold it hostage. A model this capable that no single company controls is good news for everyone who buys AI, us included.

What it does in Jyper

K3 runs the night shift: the long batches of supplier chasing, reply-processing and file updates that happen while your office sleeps. By morning its work is queued for review, with the surprises flagged. You read, adjust, approve. And the numbers in the file were computed by Jyper’s pricing engine along the way, because no model, however steady, gets to invent a price.