Three open-weight models landed in the space of a few days, and the headline numbers all point the wrong way. In Seoul, SK Telecom published A.X K2, a 688-billion-parameter model, on July 29. Two days later LG AI Research answered with K-EXAONE 2.0 at 750 billion. Both went up on Hugging Face under an Apache 2.0 licence that lets any company use or modify them commercially. The timing was not a coincidence: the two are the largest survivors of a South Korean government contest for a sovereign national model, and evaluators convene on August 8 to cut the field from four teams to three. Nothing sharpens a release schedule like an elimination round.
The parameter counts, though, describe less than they seem to. Both models use a Mixture of Experts design, which splits the network into 256 specialist modules and routes each word through only about eight of them. So a "750-billion-parameter" model actually fires roughly 37 billion parameters per token; A.X K2 lights up about 33 billion. The point of the architecture is to store a vast breadth of knowledge while paying the running cost of a much smaller model, which is why DeepSeek, Mistral and now both Korean labs have converged on it. The two teams reached the same 262,144-token context window, long enough to read several novels at once, by different routes, and split the benchmarks between them: LG's model leads on long-context retrieval, SK Telecom's on mathematics, where it cleared the gold-medal threshold on the 2025 International Mathematical Olympiad.
The most striking release of the week, though, came from a company that has been quietly falling apart. Thinking Machines, the startup founded in early 2025 by former OpenAI leaders, shipped Inkling-Small: 276 billion parameters, only 12 billion of them active, roughly a quarter the size of the original Inkling it launched two weeks earlier. It also beats that larger sibling on mathematics, reasoning and agentic coding, and sets a new open-source high on the ARC-AGI-2 benchmark. The recipe is exactly the technique now causing diplomatic friction elsewhere: use the big model as a teacher, distil its behaviour into the student, then run two weeks of reinforcement learning until the student overtakes the teacher.
What makes Inkling-Small matter is not the leaderboard, it is the hardware bill. The original Inkling needed something like two terabytes of pooled GPU memory just to load; the smaller model, with 12 billion active parameters, is what LMSYS calls the sweet spot for fine-tuning, where both lightweight LoRA and full training become "achievable" for a mid-sized team rather than only a tech giant. Customising a frontier-grade model used to be a privilege of the very largest labs. It is quietly becoming something a competent startup can do in a spare rack.
The subplot is almost operatic. Four of Thinking Machines' six founders have now left, three of them drifting back to OpenAI, including Weng Li, who departed and returned to that gravity well just days before this launch. And yet the model still shipped, on time and improved. That is the real signal in a week obsessed with size: the frontier is being rebuilt around efficiency rather than scale, the winning tricks are Mixture of Experts and distillation rather than raw parameters, and the results keep arriving as open weights that anyone can download. The race is no longer only about building the biggest brain. It is about how little of one you need to run.