prime-rl 0.11.0 ships: 1.5x faster training, new pretraining entrypoint, Mega MoE dispatcher
samsja19 · x · 2026-11-22
Prime Intellect released prime-rl 0.11.0 (285 commits, 25 contributors, with breaking changes), a major update to its open-source RL training framework.
- New pretraining entrypoint: uv run pretrain trains a model from scratch on datatrove token shards with weighted data mixes, per-document validation, and exact resume
- Mega MoE: fused expert-parallel dispatcher overlapping EP communication with expert GEMMs, plus least-loaded expert rebalancing; DeepSeek-V4-Flash SFT hits 391k tokens/s/GPU (35.8% MFU), with up to 1.67x geomean speedup from rebalancing
- Compile & EP wins: Qwen3-30B-A3B step time drops 1.067s→0.703s (1.52x); RL trainer warmup cuts cold-start first step from 637s to 61s
- Memory: async checkpointing (20GB in 2-5s), 70-75% lower activation offload overhead, optimizer offloading on by default
- Bidirectional KV transfer for disaggregated P/D; env worker pools cut peak orchestrator memory from 85.3GB to 17.5GB
- New models supported: GLM-5.3 Flash, Muse Glimmer, Nemotron 3.5 Super, Qwen3.8 Flash Next
The author also notes agents are massively accelerating development, with a noticeable slop drop after the Opus 5.5 release.
More from coding & agent
- MIT study: coding agents boost commits 240% but releases only 30% — apps ship into the void — GaryMarcus · 2026-11-22
- gemini-cli PR fixes ACP mode hanging on initialize when sandbox is enabled — Raamia · 2026-11-22
- Open-Source Free PBR Material Generator Built on Apple Depth Pro 2 — joemeno · 2026-11-22
- Open-source coding agent control plane T3 Code claims 500K+ devs orchestrating Claude Code, Codex and more — churchkey · 2026-11-22
- Adding verification and accountability to coding agents: 6 auto-checked claims, 455/455 tests pass — jasonkneen · 2026-11-22
- AI computer use mocked as 'tentacle barrel' self-driving: screenshots to GPU cluster just to move a cursor — ZeroStateReflex · 2026-11-22