Agentic AI Platform Change Log

Deployment / data residency

Prime Intellect released Prime Flash MoE, a set of open-source CUDA kernels optimized for Nvidia Blackwell GPUs to accelerate the feed-forward pass in Mixture-of-Experts (MoE) models. Integrated into the prime-rl framework, these kernels keep intermediate activations on-chip, achieving up to a 2.4x speedup over PyTorch grouped GEMM.

Infrastructure teams and developers running large-scale MoE models can reduce inference latency and hardware overhead on Nvidia B200 GPUs. By accelerating the computational pass and cutting memory traffic, this update lowers the cost of serving complex agentic models in production.

Agent capability

Prime Intellect open-sourced Prime Agent, a self-improving coding and research harness released under the MIT License. The system utilizes a Recursive Language Model framework and a Continual Harness, allowing the agent to manage prompts, sub-agents, skills, and memory as state variables that can be modified during execution.

Engineering teams and AI labs can deploy this harness to build autonomous agents capable of executing complex, multi-hour tasks like overnight refactors or kernel optimization. The open-source architecture and persistent IPython kernel integration provide a highly customizable, token-efficient alternative to proprietary agent frameworks without vendor lock-in.

Integrations

Prime Intellect integrated NVIDIA Vera CPUs, NVIDIA Dynamo inference orchestration, and NeMo Gym open-source reinforcement learning tools into its Lab platform and sandbox infrastructure. This combined hardware and software alignment optimizes continuous reinforcement learning loops for agentic AI models.

Engineering and research teams can execute post-training workflows more efficiently on Prime Intellect infrastructure. The integration enables a reported 30% increase in sandbox throughput per CPU, lowering the overall compute costs required for continuous agent refinement.