Scale you can verify
parameters: 48 layers of width 5,120, a 2,048-token context and a vocabulary of 65,536. Built from Saga 3B on 29 September; in training in the Fast Lane allocation since 8 October.
See the training record →GH200 GPUs in one training run · 32 JUPITER Booster nodes
tokens of exposure: 12 stages of 20.41 billion and 18,000 updates after the fix
on the unchanged Polish set; it was 2.4872 at the start of T9
parameters · trained from scratch on 195.8 billion tokens
AI Factory Fast Lane EHPC-AIF-2026FL01-1001 on JUPITER Booster for 3 months. Awarded to grow Saga 3B into a model of about 15 billion parameters. The growth itself, to 15.45 billion, was done on 29 September in the Development project; since 8 October the main training of that model has been running here; by 2 November twelve stages of about 20 billion tokens (T9–T20) were complete, and after the 12 November fix to how the gradient is computed the model went through 18,000 updates on the mixture with translations by 14 November. Project account: 48,198 of 50,000 GPU-h. Counted separately from the Playground and Development allocations.
Aggregate throughput of 9.0 M tok./s on 64 GPUs, relative to ideal scaling from the first point of the series (4 GPUs). This is a measurement of our workload.
Scaling chart and all points — Polish / PL ↗Validation loss on JUPITER compared with 2.9867 from the local 33M baseline.
A technical probe of the Saga 3B configuration, measuring throughput and memory, separate from the full training run.
Development Access EHPC-DEV-2026D09-196: 3,500 node-hours for 12 months. Awarded on 10 September 2026; first job on 18 September. Counted separately from the 5,000 GPU-hour Playground allocation and intended for multi-node training development.