Look at any open-weight leaderboard in mid-2026 and count the American labs near the top. You will run out of fingers on one hand before you run out of entries from Beijing and Hangzhou. The gap to the best closed frontier models, on the kind of everyday practical tasks most people actually use these systems for, has narrowed to a few percentage points. The cost gap has moved the other direction entirely: running the open models costs roughly 4 to 10 times less per token. Both trends point at the same underlying story, and it is not primarily a story about capability.

The names doing the work

Kimi K2, from Moonshot AI, is a mixture-of-experts model with 1,000 billion total parameters, of which only 32 billion activate for any given token. That efficiency ratio is the whole trick: you get the capacity of a trillion-parameter model at something close to the inference cost of a 32-billion-parameter one. K2 has become particularly known for agentic work, the kind of task where a model has to call tools, chain several actions together, and recover when the first attempt does not work. The K2.6 release earlier this year pushed that further, with noticeably better follow-through on multi-step jobs rather than just better single-turn answers.

GLM, from Zhipu AI, has climbed to GLM-5.2: 754 billion total parameters, 40 billion active per token, a context window stretching to 1 million tokens, and released under the MIT license, about as permissive as licensing gets. GLM has carved out a reputation as one of the strongest open models specifically at coding, the kind of benchmark where subtle correctness differences actually matter rather than getting lost in the noise.

Then there is gpt-oss, OpenAI’s own return to open weights in 2025 after years of keeping everything closed, shipped in 120B and 20B versions. It is a notable release less for topping charts and more for who released it: the lab most associated with the closed-model era deciding the open lane was worth re-entering.

A few more names are worth tracking without a deep dive here: Nemotron out of NVIDIA, MiniMax, and MiMo from Xiaomi. None of them lead the pack yet, but each has shipped something in the last year good enough to suggest they will not stay in the “watch” category for long.

What this closes out

This piece closes a ten-part run through the open-model landscape, and it is worth naming the throughline directly. That series opened with the distinction between open weight and open source, a distinction that still matters more than most coverage admits. It tracked a parameter-efficiency story, from dense models to mixture-of-experts architectures that get more capability per active parameter, visible again here in K2’s 32-of-1,000 and GLM’s 40-of-754 ratios. It followed a licensing story, from restrictive terms to the MIT license GLM ships under today. And it followed Meta Llama: The Family That Started the Open Wave, a wave that, as that piece covered, started almost by accident and that Meta has since been quietly stepping back from.

Picture a relay race where the runner who fired the starting gun is no longer anywhere near the lead pack, having handed off the baton and drifted toward the back while runners who joined later carry it forward. That is roughly what happened to the geography of open models over about a year. Meta opened the lane. Chinese labs are now the ones sprinting down it, which DeepSeek: The Model That Redefined Cost already previewed as the earlier example of this same cost-efficiency shift, before Kimi, GLM, and the rest turned it into the norm rather than the exception. That reversal gets far less attention in headlines than the benchmark scores do. It may be the more consequential number in this entire series.