Nvidia releases Nemotron 3.5 Lightning on Prime Intellect with support from day one

Nvidia releases Nemotron 3.5 Lightning with immediate availability on Prime Intellect, a MoE model of 30 billion parameters that activates only 3 billion per forward pass. According to the vendor, the architecture yields up to four times higher throughput and accelerates task completion by about 30 % versus earlier versions. The model is intended for post-training fine-tuning on specific domains via prime-rl and Prime Lab, the two tools Prime Intellect offers as part of its open-source stack.

The compact 30B design with 3B active experts places Lightning in a niche where relatively small models compete with larger ones on quality while cutting inference costs noticeably. The MoE (Mixture of Experts) routing sends each token to only the relevant expert subset, keeping actual computation narrow even though theoretical capacity is large. Prime Intellect stresses that the model arrives “ready for post-training”, meaning the weights are open for further optimization on dedicated data without restarting from scratch.

Prime Intellect’s offering goes beyond model hosting. The platform provides an integrated stack that includes on-demand compute, reinforcement-learning-based training environments, dedicated or serverless inference servers with native LoRA support, and large-scale sandboxes for secure code execution. Installation is performed with `pip install prime`, and the CLI manages the entire workflow—from defining an RL environment, through asynchronous distributed training, to deployment and feeding production traces back into the next training cycle.

According to Karim Atiyeh, co-CEO of Ramp, the company trained a sub-agent called Fast Ask on Prime Lab to help spreadsheet agents locate answers inside tables. He claims the result outperformed frontier models in accuracy, ran faster, and cost a fraction of the price. The message is that organizations can train a narrow model for a critical task and obtain immediate ROI instead of waiting for a better general model.

The platform’s Environment Hub aggregates more than 2,500 open-source RL environments, ranging from mini-swe-agent-plus and deepdive to text-game and science-env collections such as hud-text-2048. The verifiers library supplies modular components for building environments—datasets, reward functions, evaluation rubrics, and action tools—exposed through a `ToolEnv` interface that allows custom turn limits and tool specifications. The infrastructure supports asynchronous scaling via separate configuration files for trainer, scheduler, and inference.

Prime Intellect runs a global GPU marketplace. Spot prices for H200 start at 0.47 dollar per hour (1.70 shekel per hour), while reserved instances reach up to 4.99 dollar per B300. Liquid Reserved clusters let users request hundreds of cards from multiple vendors (from 50 providers)—for example 512 B300 SXM6 units at 5 dollar per hour per card (totaling roughly 2,560 dollar per hour, about 9,300 shekel)—with the option to sell unused capacity back to the spot market. Transparent pricing is intended to enable predictable budgeting for long-running training jobs.

In parallel with the commercial product, Prime Intellect publishes research directions: Prime Agent, built around Recursive Language Models (RLM) and Continual Harness for ongoing self-improvement, and a position paper framing recursive models as the leading paradigm for 2026. The idea is a model that writes its own training code, evaluates performance, and generates the next version in a closed loop, using production traces as input for the subsequent iteration.