Introducing Olmo-core 3: Open, scalable training infrastructure for large MoEs

Synthesis

1. 01

   Maintaining throughput while scaling expert count

   The post reports that increasing the expert pool from 8 to 128 while selecting only four experts per token kept active parameters per token roughly fixed at ~3.2B; total parameter capacity grew from 4.6B to 47B with less than 5% drop in training throughput.

   Supporting evidence 1

   Original excerpt

   In one benchmark, we increased the expert pool from 8 to 128 while still selecting only four experts per token – the small units of text a language model processes – keeping the number of active parameters per token roughly fixed at about 3.2B. Total parameter capacity grew from 4.6B to 47B, while training throughput fell by less than 5%.

   Kyle Wiggers · Paragraph 5

   Introducing Olmo-core 3: Open, scalable training infrastructure for large MoEs · Oct 1, 2026

   Context

   Olmo-core 3 is built to close that gap.

   Read in source context →

   Continue exploring

   MoE scaling efficiency →
2. 02

   2.7× throughput gain with DDP-based MoE stack

   The post reports that on eight NVIDIA B300 GPUs, Olmo-core 3’s new distributed data parallelism (DDP)-based MoE training stack achieved 52,000 tokens/sec/GPU versus 19,400 tokens/sec/GPU with the prior FSDP-based implementation — a ~2.7× throughput improvement.

   Supporting evidence 1

   Original excerpt

   In a preliminary test on eight NVIDIA B300 GPUs, a 47-billion-parameter MoE processed 52,000 tokens per second per GPU with the new stack, compared with 19,400 using our earlier implementation—about 2.7× the throughput.

   Kyle Wiggers · Paragraph 11

   Introducing Olmo-core 3: Open, scalable training infrastructure for large MoEs · Oct 1, 2026

   Context

   NVIDIA’s Megatron-Core is an established option for training large MoEs. Olmo-core 3 brings an integrated MoE training stack to the framework behind Olmo, with a redesign that improves throughput over our earlier FSDP-based implementation.

   Read in source context →

   Continue exploring

   training infrastructure performance →
3. 03

   MXFP8 improves throughput and peak active memory in a controlled B300 benchmark

   The post reports that in a controlled benchmark on four NVIDIA B300 GPUs with uniform expert load, enabling MXFP8 in performance-critical parts increased end-to-end training throughput by ~21% over BF16 baseline and reduced peak active memory from 103 GiB to 95 GiB; most gains came from feed-forward computation and inter-expert data movement, not attention alone.

   Supporting evidence 1

   Original excerpt

   With MXFP8 enabled across the parts of the system where it helped most, training throughput was about 21% higher than with BF16, the higher-precision format we used as our baseline, while peak active memory fell from 103 GiB to 95 GiB. Most of the gain came from feed-forward computation and moving data between experts rather than attention alone.

   Kyle Wiggers · Paragraph 21

   Introducing Olmo-core 3: Open, scalable training infrastructure for large MoEs · Oct 1, 2026

   Context

   We measured MXFP8’s effect on end-to-end training throughput in a controlled benchmark on four NVIDIA B300 GPUs, with work distributed uniformly across experts.

   Read in source context →

   Continue exploring

   numerical precision optimization →