52,000 against 19,400. Tokens per second on a single chip, same model, same hardware, new code. The rest of Ai2's post is how they got there.
Fact
The facts: on 1 October 2026 Ai2 (Allen Institute for AI) published Olmo-core 3, an open training system for large mixture-of-experts (MoE) models. Per Ai2, the number of experts grows from 8 to 128 at roughly 3.2 billion active parameters per token, while total capacity rises from 4.6 to 47 billion parameters with a throughput drop of under 5%. In a preliminary test on 8 NVIDIA B300 chips a 47 billion parameter model does 52,000 tokens per second per GPU against 19,400 with their earlier implementation, about 2.7 times more. The largest test is 1.2 trillion total parameters (58.36 billion active per token) on 512 GPUs, peaking at 858 TFLOP/s per chip (tests with random routing, measuring system speed, not model quality); with DeepEP v2 the system has been tried up to 2.38 trillion parameters. Released are a technical report, the code on GitHub and an interactive demo. The post does not mention trained models or checkpoints. The code in the allenai/Olmo-core repository is under the Apache 2.0 licence.
Be careful about what this is and what it is not. There is no new model. Nobody downloads a new Olmo today. This is the tool a model is made with, and it comes from a team that so far has published data, code and weights together.
This is backstage work, like at any event: power, cables and the schedule nobody in the audience sees. The network between 512 chips is exactly that kind of work. When 128 experts have to swap tokens over thousands of links, you lose time waiting, not computing. Ai2 says the jump from 8 to 128 experts costs it under 5% in speed. Their number, their tests.
The model is the stage. This is the sound and the lights.
Who is it for? Universities and small labs that have chips but not a hundred engineers to write this kind of code. Until now they chose between the closed code of the big players and their own, which does not scale. Now there is a third way, and it is free.
The test is simple. It will show when someone outside Ai2 trains a model with it and releases it. Until then it is good code with a good report.
Sources
Ai2 - Introducing Olmo-core 3: Open, scalable training infrastructure for large MoEs, 01.10.2026Ai2 - Olmo-core (GitHub, Apache 2.0), 01.10.2026The visual is generated code-art, no third-party images.