Introducing Olmo-core 3: Open, scalable training infrastructure for large MoEs | Ai2

Synthesis

1. 01

   Maintaining throughput while scaling expert count

   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%.

   Ai2 · Paragraph 4

   Introducing Olmo-core 3: Open, scalable training infrastructure for large MoEs | Ai2 · 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 over prior FSDP-based MoE stack

   On eight NVIDIA B300 GPUs, Olmo-core 3 achieved 52,000 tokens/sec/GPU for a 47B-parameter MoE, compared to 19,400 tokens/sec/GPU with the earlier FSDP-based implementation — a ~2.7× improvement in throughput.

   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.

   Ai2 · Paragraph 10

   Introducing Olmo-core 3: Open, scalable training infrastructure for large MoEs | Ai2 · 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 yields 21% higher throughput and lower memory vs BF16

   In a controlled benchmark on four NVIDIA B300 GPUs with uniform expert load, enabling MXFP8 across optimal system components increased end-to-end training throughput by ~21% versus BF16 baseline, while peak active GPU memory decreased from 103 GiB to 95 GiB.

   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.

   Ai2 · Paragraph 20

   Introducing Olmo-core 3: Open, scalable training infrastructure for large MoEs | Ai2 · 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. Most of the gain came from feed-forward computation and moving data between experts rather than attention alone.

   Read in source context →

   Continue exploring

   numerical precision optimization →