Ai2 Releases Olmo-core 3: Open MoE Training Framework Scales to 128 Experts with <5% Throughput Loss
allen_ai · x · 2026-10-01
Ai2 released Olmo-core 3, a redesigned open mixture-of-experts training system built to scale MoE training toward the trillion-parameter range. It underpins the next generation of Olmo.
- Benchmark: growing the expert pool from 8 to 128 (still 4 experts per token, 3.2B active params) raised total capacity from 4.6B to 47B with less than 5% throughput loss; the infrastructure has been benchmarked beyond 1T total parameters
- Motivation: expert routing and coordination costs erode MoE efficiency advantages at scale
- The tech report shares experiments, findings, and failures, including "token gerrymandering," where a load-balancing score could improve as workloads became less balanced
Related event: Ai2 Open-Sources Olmo-core 3, a Trillion-Parameter MoE Training Stack(5 posts)→
More from Infra
- Gemma.c: full Gemma 1/2/3 multimodal inference in a single C file, ~72 tok/s on CPU at INT8 — This-Purchase-3325 · 2026-11-16
- Sarvam AI Launches Token Factory: India-Hosted Inference for Open Models — itsOmSarraf_ · 2026-11-16
- BIS Issues Interim Final Rule Addressing Remote Access to Advanced AI Chips — apples_jimmy · 2026-11-16
- Jailbroken PS4 runs a 9B open-source model at up to 43 tok/s by talking directly to its 2013-era GPU — AIFlow_ML · 2026-11-16
- Mobile AI Factory: agents auto-build on-device SLMs with 19x less GPU time — pmttyji · 2026-11-16
- vLLM serves DeepSeek V4.1-Flash at 954K tokens/s per GPU, $38.96/GPU-hour on GB300 NVL72 — bookwormengr · 2026-11-16