Prime Intellect Launches Open Superintelligence Stack for Agent Training
willcb · x · 2026-08-21
Radhika Gaonkar announced joining Prime Intellect. The company released the "Open Superintelligence Stack," integrating compute, training, inference, and a sandbox to help users build and own their models.
Key features include:
- Lab: Turns tasks into RL environments using the Verifiers library with a CLI loop.
- Evaluations: Hosted benchmarks for 100+ open-source models with a public leaderboard.
- Hosted Training: Trains models optimized for agentic workflows on 2,500+ RL environments.
- Inference: Dedicated or serverless inference for custom models with native LoRA support.
Backed by Founders Fund, NVIDIA, Intel, and Andrej Karpathy, a case study from Ramp shows their sub-agent outperforming frontier models in accuracy and speed.
More from coding & agent
- Single binary harness orchestrates Claude, GPT, DeepSeek, and GLM simultaneously — K_Kolomeitsev · 2026-08-21
- PuppyOne Uses Filesystem as Agent State, Bypassing Custom Stores — rohanpaul_ai · 2026-08-21
- Dev argues for upfront alignment over review in AI agent workflows — mattpocockuk · 2026-08-21
- Dev builds a game in 6 minutes using ChatGPT, showcasing AI speed — aziz4ai · 2026-08-21
- Using Agent Orchestration: Easy to use but slower and less deterministic — mattpocockuk · 2026-08-21
- LLM Financial Verification Benchmark: Deterministic 100%, Live LLM 29% — MuhammadMujtaba21 · 2026-08-21