Nvidia is spending 26 billion dollars over the next 5 years will be spent on open source artificial intelligence Models according to A 2025 financial filing. In interviews with WIRED, executives confirmed this news that had not previously been reported.
A large investment can see Nvidia The evolution of a chipmaker into a true software developer is a remarkable feat. frontier lab Can compete with OpenAI You can also find out more about the following: DeepSeek. Since the models were tuned to Nvidia hardware, it’s an important strategic move.
Open source models are ones where the weights or the parameters that determine a model’s behavior are released publicly—sometimes with the details of its architecture and training. It allows users to run the models on their machines or in the cloud. Nvidia also discloses the technological innovations used in the building and training of its models. This makes it easier for entrepreneurs and researchers alike to adapt and improve upon Nvidia’s innovations.
Nvidia’s Nemotron 3 Super is the most advanced open-weight AI AI model it has ever released. It has roughly 128 billion parameters, which is a measure of model complexity. This model is similar to OpenAI GPT-OSS’s largest version. However, the company says it performs better than GPT OSS across multiple benchmarks.
Nvidia says that Nemotron 3 Super scored 37 out of 100 on the Artificial Intelligence Index. This index scores models based on 10 benchmarks. GPT-OSS scored 33—but several Chinese models scored higher. Nvidia claims that Nemotron 3 Super has been secretly tested using PinchBench. This new benchmark measures a model’s ability to use OpenClaw. Nemotron 3 Super ranks first on this test.
Nvidia used a variety of tricks to train Nemotron 3 as well. They also introduced Nvidia’s own set of tricks to help train Nemotron 3 include architectural and training techniques This will improve the model’s ability to reason, handle long contexts, and respond to reinforcement learning.
“Nvidia is taking open model development much more seriously,” Bryan Catanzaro is the Vice President of Deep Learning Research at Nvidia. “And we are making a lot of progress.”
Open Frontier
Meta is the first major AI company. release an open modelLlama will be in 2023. Mark Zuckerberg, the CEO of Facebook, recently re-started its AI efforts. signaled that It is possible that future models will not be completely open. OpenAI provides an open-weighted model. called GPT-ossThe software is not as advanced or flexible to be modified, and falls short of the most innovative proprietary products.
OpenAI’s best US models AnthropicThen, GoogleOnly the cloud, or a chat-based interface can access. The weights of many Chinese top models from DeepSeek and Alibaba to Moonshot AI and MiniMax, are freely available. In response, startups and researchers from around the globe are building upon Chinese models.
“It’s in our interest to help the ecosystem develop,” Catanzaro joined Nvidia as a member in 2011, and has been instrumental in the shift away from graphics cards used for gaming, to silicon that is AI. Nvidia launched the first Nemotron in November 2023. He adds that Nvidia recently finished pretraining a 550-billion-parameter model. The pretraining process involves feeding large amounts of data simultaneously into many specialized processors to create a complex model. Nvidia released models that are specialized in robotics, climate modeling, and protein fold.
Kari Briski is the VP for generative AI enterprise software at Nvidia. She says that the future AI models of the company will improve the chips and the datacenters they build. “We build it to stretch our systems and test not just the compute but also the storage and networking, and to kind of build out our hardware architecture roadmap,” “She says”
