Three new open-weight models released: Laguna S2.1, Inkling, and Kimi K3 join Pareto frontier

Poolside introduced Laguna S2.1, an updated version of its 118B-A8B mixture-of-experts (MoE) model, designed for pre- and post-training and optimized to run on DGX Spark systems. The company also adopted the OpenMDW license, a free license with legal safeguards tailored for AI models.

Thinking Machines debuted Inkling, a 975B-A41B multimodal MoE model supporting text, images, and audio as inputs and producing text outputs. The company simultaneously released a smaller 276B-A12B variant, positioning both as strong candidates for fine-tuning via its commercial offering, Tinker.

Moonshot AI launched Kimi K3, described as the largest open model release in some time. The model is distributed under a noncommercial license, which requires inference and fine-tuning providers to enter into a commercial agreement. Analysts note this licensing structure could influence future policy actions regarding U.S.–China AI collaboration.

Tencent released Hy3, a 295B-A21B MoE model, with improvements over its predecessor across multiple metrics. The company also switched from a custom restrictive license to Apache 2.0 for this release. The model demonstrated the ability to solve a 50-year-old math problem under controlled evaluation conditions.

DeepSeek updated its V4-Flash-0731 model, positioning it at the Pareto frontier one day after OpenAI reduced prices for its smallest model. The company has not yet updated its larger V4 model, leaving its comparative performance unspecified.

Additional releases included Meituan-LongCat 2.0, a 1.6T-parameter MoE model trained entirely on Ascend 910 accelerators, marking a notable use of non-Huawei Chinese chips for training. AMD also contributed Instella-MoE-16B-A3B-Think, a 16B-A3B MoE trained on Instinct cards, with full training and fine-tuning artifacts provided.