Latest Open Artifacts (#23): Laguna S2.1, Inkling, & Kimi K3 Show the Utility of Open Models on the Pareto Frontier
Source
Florian Brand
Author
Florian Brand
Date
Key takeaways · AI-distilled
Thinking Machines' Inkling is a 975B-A41B mixture-of-expertsA model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.Full definition → model taking text, images and audio in and text out. It isn't the strongest in its size class; it's built to be fine-tuned, feeding the company's paid tuning service Tinker.
Poolside's Laguna-S-2.1 is a 118B-A8B MoE sized to fit on a single DGX Spark desktop box, and that fit is most of why it got attentionThe mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.Full definition →. Poolside also published the full evaluation trajectories, which open releases almost never do.
OpenMDW is an Apache-2.0-style free license written specifically for model weights. MIT and Apache are software licenses, so their legal footing over weights and data is shaky; OpenMDW closes that gap. NVIDIA and Poolside both adopted it.
Tencent's Hy3 (295B-A21B MoE) dropped the restrictive custom license its predecessor shipped under and went Apache 2.0, converting a look-but-don't-touch release into one you can build a commercial product on.
Meituan's LongCat-2.0, a 1.6T-parameter MoE, was trained entirely on Huawei Ascend 910s. Chinese accelerators are normally used only for inferenceRunning a trained model to get answers — the phase where AI is actually used, as opposed to trained.Full definition →, so this is the first serious model trained end to end without NVIDIA silicon.
Terms in this piece · Glossary
mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
attention — The mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.
inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
Why it matters
Tracks fast-moving shifts in open weightsA model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.Full definition → model licensing (e.g., Apache 2.0 conversions vs. noncommercial revenue-share terms) that directly affect what engineers can legally build on and deploy.
Key quotes
“But if a US company needs a contract with Moonshot to provide the inference tokens that Kimi K3 generates, the picture looks different.”
Kevin Xu and Graham Webster
“Some of the policy tools US officials and others have debated as potential levers to restrict Chinese open model use would more clearly apply.”
Kevin Xu and Graham Webster
“This is a lot of transparency for an open model release!”
