Latest open artifacts (#23): Laguna S2.1, Inkling, & Kimi K3 show the utility of open models on the Pareto frontier

Everyone with a chart and a podcast has been calling for consolidation in AI labs for two years now. The logic seemed airtight: training costs balloon every generation, so eventually only a handful of well-funded giants could afford to play. Except that's not what's happening. More organizations than ever are training frontier-adjacent models and, notably, giving them away.

The clearest proof point is Thinking Machines, the company that shipped its first open model, Inkling, this week. When Thinking Machines launched back in February 2025, almost nobody pegged them as an open-weights outfit. Now their Tinker fine-tuning service is reportedly pulling in hundreds of millions annually, and Inkling — a 975-billion-parameter mixture-of-experts model with a 41-billion active parameter footprint, handling text, image, and audio input — puts them ahead of Nvidia's Nemotron and Arcee's Trilogy among US-built open models. It's not the strongest model in its weight class globally, but it's built to be a fine-tuning base, which is exactly the business Thinking Machines is running.

China isn't slowing down either. Tencent's Hy3, a 295B-A21B MoE, upgraded across the board from its predecessor and switched from a restrictive custom license to Apache 2.0 — a meaningful shift in openness that's easy to overlook next to the model's claim of solving a 50-year-old math problem (a claim that leans on a custom evaluation harness, so take it with some salt). Poolside, a company that's now shown up in three consecutive monthly roundups, released Laguna-S-2.1, a 118B-A8B model that fits on a single DGX Spark box, adopting the OpenMDW license and publishing full evaluation trajectories — a level of transparency that's rare even among open-model releases.

Then there's Kimi K3 from Moonshot AI, arguably the biggest open release in months. It ships under a noncommercial license that forces any inference or fine-tuning provider into a direct commercial agreement with Moonshot. Kevin Xu and Graham Webster have pointed out this structure isn't just a licensing quirk — it creates a contractual dependency that could become a lever for US policy action against American companies doing business with Chinese AI firms, since those firms would need an actual agreement with Moonshot to serve K3's tokens.

DeepSeek, meanwhile, quietly dropped V4-Flash-0731 one day after OpenAI cut prices on its smallest model by 80%, and the update reportedly beats Luna on the performance-per-parameter frontier. The full V4 Pro model hasn't been refreshed yet, so where the flagship lands is still an open question. Add in Meituan's LongCat-2.0, an eyebrow-raising 1.6-trillion-parameter MoE trained entirely on Huawei's Ascend 910 chips rather than Nvidia silicon, plus smaller entries from Motif, Swiss AI, and AMD, and you get a picture of an ecosystem that's getting wider, not narrower.