Poolside, Thinking Machines, Moonshot ship open-weight models

The prediction from a couple of years back was that AI would consolidate. Training costs were climbing by orders of magnitude every year, so the natural end state was a handful of labs at the frontier and everyone else buying tokens from them. The latest roundup from the Interconnects newsletter makes that story harder to hold, cataloguing a batch of open weight releases that landed together.

Poolside's Laguna S2.1 is a 118B mixture-of-experts model with 8B active parameters, small enough to fit on a DGX Spark, released under the OpenMDW license, which the piece describes as an "Apache 2.0-like free license but has better legal backing for AI models specifically." The company also published its evaluation trajectories, which the author calls "a lot of transparency for an open model release." Thinking Machines' first model, Inkling, is a 975B-A41B multimodal MoE that supports text, images, and audio as inputs, and it is positioned as a base for the company's Tinker fine-tuning product. Moonshot's Kimi K3, described as "the biggest open model release in some time," lands under a noncommercial license that requires inference and fine-tuning providers to enter into a commercial agreement.

What matters is that these are landing at very different scales, from a workstation-sized MoE up through a large frontier-adjacent one, from several different organizations across the US and China. If your planning assumed that only two or three closed labs would be worth talking to, that assumption is aging faster than expected. There are nearby data points in the same roundup: a DeepSeek V4 Flash update arrived, per the piece, "one day after OpenAI has dropped the prices of their smallest model by 80%," Tencent moved Hy3 to Apache 2.0, and Meituan's LongCat-2.0 was, in the newsletter's words, "trained entirely on Ascend 910s, making it the first non-Huawei, non-toy model trained entirely on Chinese accelerators."

The honest caveat is that this roundup does not give you head-to-head benchmark numbers, so any claim about exactly where each model sits on the capability curve has to be taken as the author's framing rather than settled fact. The Kimi K3 licensing arrangement is also real friction for US buyers, and the exact terms of that required commercial agreement are not spelled out. What the reporting does make clear is the shape of the market: token demand stays high enough that fine-tuning services and inference businesses remain viable, and open weights keep arriving at scales that were supposed to be uneconomic to release. For teams building on top of models, the interesting move is planning for a world where the base model you fine-tune next quarter probably is not the same one you fine-tuned this quarter, and probably was not trained by the vendor you expected.