The open model world now runs on three kinds of makers

A year ago the list of serious open-weight model makers was short and mostly Chinese. That has changed. The latest open artifacts roundup on Interconnects argues that the open model world is now split across three kinds of organizations, and the mix is what makes it durable. First are the pure model makers, companies like DeepSeek, Poolside, Zyphra, Cohere, and Mistral whose product is the model itself. Second is Big Tech, where Alibaba, Google, and NVIDIA each open-weight for different reasons: Alibaba uses Qwen to sell its closed models, while NVIDIA simply wants more GPU demand. Third are product companies such as JetBrains and Zed that release narrow models tuned to their tools, with no revenue to cannibalize.

The concrete releases carry the point. NVIDIA put out Nemotron-3-Ultra-550B under the new OpenMDW license, which is written specifically for model weights. Cohere shipped Command A+, a 218B mixture-of-experts model, under Apache 2.0, moving away from its earlier non-commercial terms. Zyphra released ZAYA1-74B trained on AMD hardware, and Poolside opened Laguna-M.1, also Apache 2.0. GLM-5.2 gets called out as good enough for everyday work with little gap to the leading closed models. The piece reads the spread as a reason not to restrict open models, since doing so would hand power to a few actors.

Why it matters

If you pick models for a product, the useful signal here is the license drift toward plain Apache 2.0 and weight-specific terms, which lowers the legal friction of shipping open weights. The field is wide enough now that you can choose on fit and cost rather than settling for whoever released last.