Latest open artifacts (#22): Zyphra, Cohere, and Poolside are expanding the breadth of the ecosystem
- Source
- interconnects.ai
- Author
- Florian Brand
- Date
Context
The open-model releases in this roundup give teams more options for running models under their own control. Cohere's Command A+ model card and Poolside's Laguna M.1 card both provide downloadable weights under Apache 2.0.
Deployment remains a separate engineering commitment. The Interconnects roundup notes that earlier Command releases used a non-commercial license, and that Command A+ becomes usable on a single B200 accelerator when run at four-bit precision. Lower precision reduces the memory a model needs, so the chosen numerical format shapes the hardware a team must provide.
- Big Tech releases are loss leaders with a specific job: Alibaba uses Qwen to upsell its closed models, and NVIDIA benefits because a busy open ecosystem sells more GPUs. That's a sturdier motive than Meta's Llama-era one, which didn't hold.
- Companies like JetBrains, Zed, Krea and Photoroom sell products, not , so they train small models tuned to one product need and open the weights. Giving away the weights costs them nothing because the product, not the model, is the business.
- Cohere's Command A+ is a 218B-A25B that runs on a single B200 at 4-bit precision, and it shipped under Apache 2.0 after earlier versions in the series were noncommercial-only.
- GLM-5.2 is good enough for everyday work, close to the best closed models, yet its download numbers tracked GLM-5's almost exactly. Capability gains don't automatically translate into adoption of an open release.
- NVIDIA's Nemotron-Labs-Diffusion-14B runs in three modes from one set of weights: autoregressive, diffusion, and self-speculation, each suited to a different workload.
- open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
- token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
- mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
“Attempts to slow or ban this ecosystem are not only futile, as the history of tech-related bans has shown, but also unsafe and anti-freedom .”
Checking sign-in…
Loading comments…