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

A new open-artifacts roundup describes a wave of openly released models, arguing that predictions of lab consolidation have not materialized as more organizations train and release strong models.

Thinking Machines released Inkling, described as a 975B-A41B multimodal MoE accepting text, images and audio as inputs and producing text output, plus a smaller 276B-A12B version. The roundup states it is positioned as a fine-tuning base, for example through the company's Tinker offering, and says the company's open model fine-tuning service is making hundreds of millions in revenue per year.

Tencent's Hy3 is described as a 295B-A21B MoE that improves over its predecessor across all metrics and switches from a custom restrictive license to Apache 2. The roundup says it proved a 50 year old math problem with a dedicated harness and Sol as a judge, adding it is unclear how important the judge was.

Poolside released Laguna-S-2.1, a newly pre- and post-trained version of its 118B-A8B MoE that fits on a DGX Spark, under the OpenMDW license, and updated its 33B-A3B Laguna-XS-2.1. DeepSeek released V4-Flash-0731, described as beating Luna at the Pareto frontier, one day after OpenAI cut prices of its smallest model by 80 percent.

Kimi K3 was released under a noncommercial license requiring inference and fine-tuning providers to enter commercial agreements. Kevin Xu and Graham Webster argue such licenses could enable future government action against US entities doing business with Chinese AI companies.

Other releases include Meituan's LongCat-2.0, a 1.6T-parameter MoE trained entirely on Ascend 910s, described as the first non-Huawei, non-toy model trained entirely on Chinese accelerators, and Motif-Technologies' Motif-3-Beta, a 314B-A13B MoE preview with GDLA and mHC architectural innovations.