From 46M to 325M: An Inference NPU Now Pretrains a Language Model at GPU Quality
aimachinelearningedgeainpu
TL;DR
Two weeks ago we wrote about VKORE-46M, a small model pretrained on an inference-only NPU. VIDRAFT has now scaled the same idea seven times: VKORE-325M-NPU-Base, a 325.6M-parameter language model pretrained from random initialization on a Rebellions ATOM-Max, a commercial inference NPU. A twin model trained on an NVIDIA B200 under identical settings is the reference. On 1,024 windows of unseen text, the NPU model scores 0.7993 and the GPU model 0.7994.
Why this matters for edge and on-site AI
Inference NPUs are already installed in data centers and on premises to serve models. If the same chip can also train and fine-tune, a site can adapt models to its own data without shipping that data to a GPU cloud and without buying a second kind of hardware.
| VKORE-46M (Oct 2026) | VKORE-325M | |
|---|---|---|
| Parameters | 46M | 325.6M |
| Start | random init | random init |
| Hardware | inference NPU | Rebellions ATOM-Max (inference NPU) |
| Reference | none | identical B200 twin |
| Unseen-text loss | n/a | 0.7993 (NPU) vs 0.7994 (GPU) |
How the comparison was made
Everything was held fixed except the hardware: same initialization, data, batch and learning-rate schedule. The run used 5 billion byte tokens of FineWeb-Edu over 76,292 steps (about four days) and finished without divergence. The training curves overlap end to end. The chip alone does not train; VIDRAFT's own software stack makes it possible. Both models' weights and the evaluation script are public.
FAQ
Is the NPU model as good as the GPU model? On unseen FineWeb-Edu text, yes: 0.7993 vs 0.7994 loss.
Was any pretrained checkpoint used? No. Both started from random weights.
Can I check the result? Yes. The weights for both models and the evaluation script are public, so anyone can re-run the quality comparison. The training method itself is not published (patent pending).
Is this a large model? Not yet. 325M is a research-scale model; the point is that the hardware can do the job.