01 — Measurements

Scale you can verify

These figures describe our training and its configuration on JUPITER Booster. They are not a performance assessment of JUPITER or its host.
Report timestamp
Model in training · Saga 15.45B
15.45 B

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.

Allocation
50,000 GPU-h
Used
48,198 GPU-h
See the training record →
Training scale
128

GH200 GPUs in one training run · 32 JUPITER Booster nodes

Training in Fast Lane
263.8 B

tokens of exposure: 12 stages of 20.41 billion and 18,000 updates after the fix

Validation loss · end of T20
2.4273

on the unchanged Polish set; it was 2.4872 at the start of T9

Base model · Saga 3B
3 B

parameters · trained from scratch on 195.8 billion tokens

Third EuroHPC allocation · Fast Lane
50,000 GPU-h

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.

Awarded on 25 September 2026Training since 8 October · until 31 December 2026 · no extensions
Earlier measurements · September 2026
Phase A2 · scaling efficiency at 64 GPUs
89.5%

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 ↗
A1 · Reproducing the reference
2.9847

Validation loss on JUPITER compared with 2.9867 from the local 33M baseline.

Difference −0.002013 min 11 s
Probe · 3B model
13,883 tok./s per GPU

A technical probe of the Saga 3B configuration, measuring throughput and memory, separate from the full training run.

MFU · 50.2%76.1 GiB per GPU
Second EuroHPC allocation
14,000 GPU-h

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.

Editorial snapshot: 7 September 2026 · 19:24:40 Europe/Warsaw. Reading the current published report requires JavaScript.

Task history readout: 7 September 2026, 19:24:40 (Warsaw). Measurements and the plan retain their FAKTY source date: 7 September 2026, 17:46:39 (Warsaw).