Descriptions:
Two Minute Papers reviews NVIDIA’s Neotron 3 Ultra, the company’s newest free and fully open AI model, following several days of firsthand experimentation. The model ships with 550 billion total parameters — roughly 10% active per token via a mixture-of-experts architecture — a 1 million token context window, and Mamba memory layers that compress conversation context efficiently rather than rereading the full history on each pass. Low-precision NVFP4 numerics and speculative multi-head token drafting round out the architectural highlights.
The reviewer’s coding experiments were a mixed result. Requests for a light simulation program and a real-time strategy game both produced black screens or near-empty output, and the model generated over 1,000 lines of code for tasks a 250-line handwritten solution handles cleanly. DeepSeek R4 Flash handled the same prompts more capably. However, Neotron 3 Ultra performed excellently on terminal tasks, quick experiments, file organization, and general-purpose assistance — and its speed is a genuine standout. NVIDIA made iterative improvements after the reviewer reported issues during the early-access period.
A highlight of the video is a careful licensing breakdown. Neotron 3 Ultra ships under the OpenMDW license — essentially Apache 2.0 tailored for machine learning weights — which the reviewer rates 9 out of 10 for openness, a major step up from NVIDIA’s previous proprietary license. Weights, training data, and research paper are all being released. The reviewer’s practical conclusion: Neotron 3 Ultra fits best as one slot in a multi-model roster, paired with a vision-capable model like Gemma 4 for tasks it cannot handle, and is best run on cloud infrastructure like Lambda given its 550-billion-parameter size requirement.
📺 Source: Two Minute Papers · Published June 14, 2026
🏷️ Format: Review
