NVIDIA opens 100+ frontier models free

NVIDIA is not just selling chips anymore. It is trying to become the default place where developers discover, test, and then deploy frontier AI models. That is the real story behind its Build platform now showing more than 100 models, with a big chunk available through free serverless APIs and many others packaged for NVIDIA NIM deployment on NVIDIA hardware. (build.nvidia.com) ### What actually opened up? The short version is simple — NVIDIA’s model catalog has gotten big enough to matter. The Build site now lists 160 models, with filters showing 50 “Free Endpoint” models and 113 downloadable deployments. Those are not just NVIDIA’s own models. They include third-party names developers already care about, like DeepSeek, MiniMax, Google, Mistral, and Z.ai. (buil([build.nvidia.com)Which models make this feel “frontier”? Two examples tell the story. DeepSeek V3.2 is listed as a free endpoint and described by NVIDIA as a state-of-the-art 685B reasoning model with sparse attention, long context, and agentic tools. MiniMax M2.7 is also listed as a free endpoint, and NVIDIA describes it as a 230B-parameter model built for coding, reasoning, and office tasks. (build. ([build.nvidia.com)g)) ### Is this really free? For development, yes — but not in the “NVIDIA is becoming a charity” sense. The site explicitly pitches “free serverless APIs for development,” while steering users toward DGX Cloud and self-hosting for real deployment. MiniMax’s page also says usage is logged for security and governed by NVIDIA’s API trial terms. So the free tier is best understood as a try-before-you-scale layer, not a permanent blank check. (build.nvidia.com) ### Why does NVIDIA want developers here first? Because the easiest way to win enterprise AI is to own the path from prototype to production. NVIDIA NIM is the bridge. Developers can test a model through a hosted API, then move to a downloadable NIM microservice for self-hosting on their own NVIDIA GPU setup. NVIDIA’s pitch is basically: same model family, standard APIs, optimized inference stack, fewer deployment headaches. (developer.nvidia.com) ### What is NIM really selling? Convenience and performance. NIM packages inference as GPU-accelerated microservices, with engines built on TensorRT, TensorRT-LLM, vLLM, SGLang, and related tooling. NVIDIA keeps emphasizing low latency, high throughput, and the ability to run the same approach across cloud, data center, and RTX workstations. That matters to enterprises be(developer.nvidia.com) scaling usually become the bigger headache. (developer.nvidia.com) ### Why do free models help NVIDIA’s moat? Because developers do not choose infrastructure in one step. They choose a playground first. If the playground already has the models they want, plus an OpenAI-style API experience, plus a direct path into NVIDIA deployment tooling, the switching cost shows up later. Not as a lock on day one — more like a groove. Teams build demo(developer.nvidia.com) and which stack are already validated. That last part is an inference from NVIDIA’s product flow, but the funnel is pretty visible on the site. (build.nvidia.com) ### Why now? Because the model layer is getting commoditized fast. If every lab is shipping strong open or semi-open models, the money moves toward distribution, inference, and enterprise deployment. NVIDIA already dominates the hardware side. What it is doing here is extending that advantage upward — from chips into the developer surface where model experimentation starts. (build.nvidia. ([build.nvidia.com)ne? This is less about generosity than positioning. NVIDIA is making frontier models easy to try so that NIM becomes the obvious way to run them when a project turns serious. (build.nvidia.com)