INTELLECT-2 Revisited: The Open Recipe Behind the First Globally Decentralized 32B RL Training Run
willcb · x · 2026-10-14
willcb points out that a recent approach mirrors Prime Intellect's INTELLECT-2 recipe from May 2025 — the first globally distributed RL training run of a 32B reasoning model, fully open-sourced.
Highlights from the paper:
- Fully asynchronous RL across a dynamic, heterogeneous, permissionless swarm of compute contributors instead of centralized training.
- Purpose-built infrastructure: PRIME-RL framework, TOPLOC for verifying rollouts from untrusted inference workers, and SHARDCAST for broadcasting policy weights to inference nodes.
- GRPO recipe modifications plus data filtering were key to training stability; the model beat QwQ-32B, then the 32B SOTA.
A complete, reproducible reference for decentralized training and permissionless compute networks.
Related event: INTELLECT-2 Decentralized RL Training Recipe Resurfaces(2 posts)→
More from Infra
- Mesa merges YOLOX with PyTorch support in Torx NPU driver — Fcking_Chuck · 2026-10-14
- Hugging Face launches Kernel Hub to distribute compute kernels as first-class artifacts — ariG23498 · 2026-10-14
- Benchmarking disaggregated LLM serving: KV cache transfer and energy costs can erase the gains — burkov · 2026-10-14
- jitllm moves a running LLM's layers between GPUs mid-reply in 8ms with zero recompute — web3samy · 2026-10-14
- Micron hikes 2027 HBM prices as memory nears 23% of PC build cost — tengyanAI · 2026-10-14
- Beff claims TSUs paired with GPUs will deliver 100x perf-per-watt via disaggregated inference — beffjezos · 2026-10-14