NVIDIA Adopts the OpenMDW License with Nemotron-3 Ultra — The Open Model Ecosystem Keeps Diversifying

NVIDIA's newly announced Nemotron-3 Ultra-550B delivers faster inference than other models in its class, thanks to its LatentMoE architecture. The biggest headline here is that NVIDIA has dropped its previous custom license in favor of **OpenMDW**, a license purpose-built for model weights. That's one answer to a longstanding problem: MIT and Apache are software licenses and don't really fit model weights. A large portion of the training data has been released as open source as well.

Meanwhile, the latest Interconnects report (Artifacts 22) notes that the open model ecosystem is diversifying. A year ago it was dominated by Chinese companies, but today a much wider range of players have entered the space:

Pure-play model makers** (DeepSeek, Zhipu, Poolside, Zyphra, and others)

Big tech** (Alibaba's Qwen, Google's Gemma, NVIDIA)

Product companies** (JetBrains, Zed, Krea, Photoroom, and others)

Cohere has also released its flagship Command A+ under Apache 2.0, a shift away from its previous non-commercial license, and Poolside has declared that 'open weights will be the default,' committing to future releases.

New Model Release Highlights

Kimi-K2.7-Code** (Moonshot): an update focused on token efficiency

Step-3.7-Flash** (StepFun): particularly strong performance in math

Nemotron-Labs-Diffusion-14B** (NVIDIA): an experimental model that can switch between three modes — autoregressive, diffusion, and self-reasoning

Anthropic CEO Amodei Branded Open Source's 'Public Enemy No. 1'

Comments about open source from Anthropic CEO Dario Amodei have ignited a fierce debate in Reddit's LocalLLaMA community.

Amodei reportedly argued that 'with open source you can see the source, but not what's inside the model' — to which community members countered that 'that is exactly the definition of open weights.' Pointing to examples like Nemotron3 Ultra, which publishes everything — the data, the training scripts, and the model itself — they argue such releases are at least more transparent than closed models.

His claim that 'in the end you'll need to host it in the cloud' has also drawn criticism for ignoring the existence of smaller models like Qwen 27B and compact MoE models. Some are even questioning whether he has ever actually tried running an open model himself.

Mass AI-Powered Exam Cheating Uncovered at Brown University

According to a report by the Spanish newspaper El Pais, a large number of AI-assisted cheating cases on exams have come to light at Brown University, prompting professors to issue a statement. They warn that academic integrity is in crisis, and the episode is once again raising questions about where to draw the line on AI use in higher education.

The story has also gained traction on Hacker News, fueling a broader discussion that as AI tools spread, educational institutions may need to rethink how they assess students in the first place.

Analysis of 1 Million LLM API Calls: 62% Used the 'Wrong Model'

Tokonomics tracked 1 million LLM API calls and found that 62% of them used the 'wrong model' — cases where a task was handled by an overpowered (and overpriced) model, or conversely by one that wasn't capable enough.

The findings lay bare how many organizations never optimize model selection and simply keep using the default, racking up unnecessary costs. Model routing optimization is emerging as a key challenge for running LLMs in production.

Anthropic Announces Self-Hosted Sandboxes for Claude Managed Agents

According to a Qiita article, Anthropic announced at 'Code with Claude London' in May 2026 that it has added **self-hosted sandboxes** to Claude Managed Agents. The feature enables tool execution on a company's own infrastructure, and is expected to lower the barrier to agent adoption for enterprises with strict security requirements.

---