Disclosure: the work described here is by VIDRAFT, an affiliate of Ginigen AI.
TL;DR
Edge and on-premise deployments usually split hardware into two worlds: GPUs to train, cheaper accelerators to serve. A new result from VIDRAFT narrows that split. They pretrained a 325M parameter language model from random initialization on the Rebellions ATOM-Max, a commercial inference NPU released in 2025, and trained the same model on a B200 GPU for comparison. On text neither model saw in training, the losses are 0.7993 (NPU) and 0.7994 (GPU).
Why should edge and on-prem teams care?
Most organizations running AI on premise already have inference accelerators in the rack. They buy them because they are cheaper per query and draw less power than training GPUs. When a model needs adapting to local data, the usual answer is to rent GPU time or ship data out.
If the inference hardware can train, that changes the economics and the data story. Adaptation and even pretraining can happen on hardware the team already owns, inside its own network.
What was actually shown?
| Item | Value |
|---|---|
| Model | VKORE-325M-NPU-Base, 325.6M parameters |
| Start | random initialization, no pretrained weights |
| Data | 5.0B byte tokens of FineWeb-Edu |
| Run | 76,292 steps, about four days, no divergence |
| Hardware | Rebellions ATOM-Max NPU (2025) |
| Reference | the identical model on a B200 GPU |
The two training curves overlap for the whole run, and both runs follow the same learning-rate schedule.
Is the NPU-trained model really as good?
On 1,024 windows from a FineWeb-Edu shard that was never in the training corpus:
| Model | Loss | Bits per byte |
|---|---|---|
| Rebellions ATOM-Max NPU | 0.7993 | 1.1531 |
| B200 GPU reference | 0.7994 | 1.1533 |
Window by window, the difference is -0.0001 with a standard error of 0.0007. That is a tie.
How far has this line come?
What is the catch?
Two honest limits. First, the hardware alone does not do this: the ATOM-Max and its standard software are built for inference, and training depends on VIDRAFT's own training stack and stability technique, which are under patent application. Buying the chip does not by itself enable training. Second, 325M on 5B tokens is a research-scale base model, not a production assistant. The result is about what the hardware class can do.
FAQ
Which NPU was used? The Rebellions ATOM-Max, a commercial inference NPU released in 2025.
Can I verify the numbers? Yes. Both weights and evaluate_unseen.py are in FINAL-Bench/VKORE-325M-NPU-Base.
Does this mean GPUs are unnecessary for training? No. It means inference NPUs can carry real training runs when the software exists. Scale, speed and cost per run are separate questions.
What does it mean for on-device AI? Models can be adapted close to where they run, on hardware already deployed, without moving data to a GPU cloud.
Top comments (0)