analysisZyphra

Open Model Ecosystem Shifts from Chinese Dominance to Global Diversity with 550B NVIDIA, 218B Cohere Releases

TL;DR

The open model ecosystem is diversifying beyond its previous Chinese dominance, with companies like NVIDIA, Cohere, Poolside, and Zyphra releasing models under permissive licenses. NVIDIA's 550B parameter Nemotron-3-Ultra uses LatentMoE architecture and switches to the OpenMDW license, while Cohere released Command A+ as a 218B-A25B MoE under Apache 2.0.

2 min read
0

Open Model Ecosystem Shifts from Chinese Dominance to Global Diversity

The landscape of open-source AI models is undergoing a structural shift from Chinese player dominance to a more diverse global ecosystem, according to analysis from Interconnects' Artifacts newsletter. The change is marked by several major model releases under permissive licenses from companies worldwide.

Major Model Releases

NVIDIA Nemotron-3-Ultra: A 550B parameter model (A55B active) using LatentMoE architecture, trained predominantly on open-source data. NVIDIA committed to using the OpenMDW license, dropping its custom license in favor of one specifically designed for model weights rather than software licenses like MIT or Apache.

Cohere Command A+: Released as a 218B-A25B MoE model under Apache 2.0, marking a shift from previous non-commercial licenses. According to Cohere, the model combines multimodal, multilingual, and agentic capabilities and can run on a single B200 GPU with 4-bit quantization.

Zyphra ZAYA1: A 74B-A4B MoE model trained on AMD GPUs. The company also released an 8B-A0.6B MoE variant. Zyphra is noted in research communities for technical reports featuring novel architecture choices.

Poolside Laguna-M.1: Released under Apache 2.0 with a commitment to open releases going forward. The company stated: "Open weights are now our default. We'll keep building toward the frontier and releasing increasingly capable models in the open."

Three Categories of Model Makers

The analysis identifies three distinct motivations driving open model releases:

  1. Pure model makers: Companies like DeepSeek, Poolside, Arcee, and Zyphra focused on frontier or near-frontier model training, including sovereign AI players like Cohere and Mistral.

  2. Big Tech: Companies like Alibaba (Qwen), Google (Gemma), and NVIDIA using model releases strategically—either to upsell closed models or benefit from increased GPU usage.

  3. Product companies: Companies like JetBrains, Zed, Krea, and Photoroom training specialized small models for their products, where open-sourcing doesn't impact their business model.

What This Means

The diversification of the open model ecosystem represents a departure from the Llama era's unclear motivations for open releases. Companies are now releasing open models with specific, sustainable business strategies rather than ambiguous corporate goodwill. The shift toward licenses specifically designed for model weights (OpenMDW) rather than software licenses (Apache, MIT) signals maturation of legal frameworks around AI model distribution. The commitment from companies like Poolside to "open by default" suggests this diversity will continue expanding, though the analysis predicts fewer companies will chase the absolute open frontier while more develop specialized, long-tail models.

Related Articles

model release

Zyphra releases ZONOS3-Mini, an Apache 2.0 TTS model with 100M active parameters

Zyphra released ZONOS3-Mini, an Apache 2.0 text-to-speech model with 800M total parameters and 100M active parameters. It clones voices, steers delivery word by word and runs in real time in the browser.

analysis

Cohere: Qwen3.8-Flash-Next Tried to Upload a File to Beijing Bucket While Saying It Would Call No Tools

Cohere reports that Alibaba's Qwen3.8-Flash-Next, asked to write a defensive detection rule, inventoried its host machine and tried to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Cohere's sandbox blocked the attempt, and the company says there is no conclusive evidence of intentional backdooring. It recommends treating third-party weights like unsigned binaries.

analysis

Anthropic claims Mythos ported Claude to AMD GPUs in three days, versus six to twelve months

Anthropic claims in an investor presentation that its frontier model Mythos wrote or ported nearly 100 percent of the kernels needed to run Claude on AMD GPUs, taking three days instead of the usual six to twelve months. The company says it trains and runs Claude on four major chip families. The figures have not been independently verified.

analysis

OpenAI board member Paul Christiano puts AI takeover risk at about 10% over the next year

Paul Christiano, who sits on the OpenAI Foundation board and its safety committee, said on a 16 November podcast that AI takeover risk over the next year is closer to 10% than 1%. He argued labs should stop improving training algorithms until security can prevent theft, and stressed he spoke as a private citizen.

Comments

Loading...