Open model releases diversify with Zyphra, Cohere, and Poolside additions

The latest analysis from Interconnects highlights a trend toward diversification in the open model ecosystem, with contributions from a growing number of organizations worldwide. While a year ago the landscape was dominated by a small set of players, often based in China, the current environment features a wider array of niche companies releasing models across various categories. These include 'pure' model makers focused on frontier or near-frontier capabilities, such as Poolside, Arcee, Zyphra, Cohere, Sovereign, Mistral, and Trillion Labs, as well as sovereign AI players like Cohere and Mistral.

Big Tech companies are also participating, though their motivations differ. For example, NVIDIA’s open releases, such as the Nemotron-3-Ultra-550B-A55B-BF16, are positioned to drive GPU usage and ecosystem growth, while Alibaba’s Qwen uses model releases to upsell its closed offerings. Product companies like JetBrains and Zed release specialized, small models to avoid dependency on closed systems, often open-sourcing weights without harming their business models.

New model releases underscore this diversity. NVIDIA’s Nemotron-3-Ultra-550B-A55B-BF16 employs LatentMoE for faster inference and is released under the OpenMDW license, which is tailored for model weights. Cohere’s Command A+ 218B-A25B MoE combines multi-modal, multi-lingual, and agentic capabilities and is released under Apache 2.0, a shift from previous non-commercial licenses. Zyphra, known for training on AMD GPUs, released the ZAYA1-74B-preview, a 74B-A4B MoE model, and an 8B-A0.6B MoE model, both flagships in their lineup.

Poolside also released its flagship Laguna-M.1 under Apache 2.0 and committed to open releases as its default going forward. The analysis notes that GLM-5.2 from zai-org continues to demonstrate usability comparable to top closed models, despite modest download numbers. MoonshotAI’s Kimi-K2.7-Code update emphasizes token efficiency, while StepFun’s Step-3.7-Flash update targets math performance.

The report argues that attempts to restrict or ban open model development would be counterproductive, concentrating AI development among a few entities and undermining broader adoption. It frames open model development as a strength of the ecosystem, with model releases often reusing training methods, architectures, and data from other open releases.