Latest open artifacts (#23): Laguna S2.1, Inkling, & Kimi K3

Despite predictions that soaring training costs would force artificial intelligence labs to consolidate, the industry is seeing more companies invest heavily and release strong open models. Organizations realize that generating tokens is a viable path to long-term value, leading to a crowded ecosystem driven by both US startups and sustained output from Chinese labs.

Recent notable releases include Inkling by Thinking Machines, a large multimodal mixture-of-experts model designed as a base for fine-tuning. Tencent released Hy3 under an Apache 2.0 license, which notably helped prove a fifty-year-old math problem using a dedicated harness. Poolside introduced Laguna S2.1 and XS 2.1, models that fit on a DGX Spark and utilize the AI-specific OpenMDW license alongside transparent evaluation data. DeepSeek updated its V4 Flash model to beat competing Pareto frontier metrics, while Moonshot AI released Kimi K3 under a noncommercial license that requires commercial agreements for inference and fine-tuning.

Additional releases highlight expanding hardware and regional developments. Meituan introduced LongCat 2.0, a 1.6-trillion-parameter model trained entirely on Chinese Ascend accelerators rather than Huawei or Western hardware. Motif Technologies previewed Motif 3 Beta with architectural innovations, Swiss AI continued pre-training Apertus 1.5 with additional tokens, and AMD released Instella-MoE trained on its own Instinct cards.