Latest open artifacts (#22): Zyphra, Cohere, and Poolside are expanding the breadth of the ecosystem

Interconnects published the 22nd edition of its "Latest open artifacts" roundup, arguing that the open model ecosystem is becoming more diverse, with an increasing number of organizations releasing a wide range of models. It notes that a year ago the open model landscape was dominated by a handful of Chinese players, a pattern that has shifted toward more niche companies worldwide.

The roundup groups makers into three broad categories: "pure" model makers whose stated goal is training frontier or near-frontier models, including DeepSeek, Zhipu, Minimax, Poolside, Arcee, Zyphra, and sovereign AI players such as Cohere, Sovereign, Mistral, and Trillion Labs; Big Tech, including Alibaba's Qwen, Google's Gemma, and to some extent NVIDIA; and product companies such as JetBrains, Zed, Krea, and Photoroom that train specialized small models for their products.

Featured releases include NVIDIA's Nemotron-3-Ultra-550B-A55B-BF16, which uses LatentMoE and is released under the OpenMDW license, replacing NVIDIA's custom license. CohereLabs released its flagship command-a-plus-05-2026-bf16 under Apache 2.0, a change from the non-commercial license used for previous iterations; it is described as a 218B-A25B MoE combining multi-modal, multi-lingual and agentic capabilities, usable with a single B200 at 4-bit.

Z.ai's GLM-5.2 is called the biggest story of the roundup, with raw download numbers roughly in line with GLM-5 after release. Zyphra released ZAYA1-74B-preview, including a 74B-A4B MoE and an 8B-A0.6B MoE, alongside a tech report. Poolside released its flagship Laguna-M.1 under Apache 2.0 and said open weights are now its default. Moonshot AI released Kimi-K2.7-Code, an update focused on token efficiency, and StepFun released Step-3.7-Flash.

Interconnects states that open model development is not driven by a single type of actor or motivation, calling this diversity a strength of the open ecosystem, visible in tech reports that reuse training methods, architecture choices and data from other open releases. It argues attempts to slow or ban the ecosystem are futile, unsafe and anti-freedom.