VIDRAFT Trains a 46M-Parameter Model Entirely on Inference-Only NPUs — No CUDA Required

TL;DR: VIDRAFT has pre-trained VKORE-46M-NPU, a 46-million-parameter language model, using only inference-oriented NPU hardware — completely bypassing NVIDIA GPUs and the CUDA stack. This proof-of-concept directly challenges the assumption that backpropagation requires high-end GPU memory hierarchies, and signals a potential path toward hardware-agnostic ML development that could meaningfully lower the cost of AI experimentation.

What it is

VKORE-46M-NPU is a 46-million-parameter neural network model developed by VIDRAFT, the Korean Pre-AGI AI startup. Its defining characteristic is not the parameter count — it is the hardware it was trained on.

Rather than relying on NVIDIA GPUs and the CUDA ecosystem that underpins virtually all large-scale model pre-training today, VIDRAFT completed the full pre-training run on inference-only NPU hardware — silicon that is typically designed for low-power, edge, or mobile inference workloads, not backpropagation.

Key facts from the source:

• Model name: VKORE-46M-NPU
• Parameter count: 46 million
• Training hardware: Inference-oriented NPUs (no high-end GPU involvement reported)
• Training type: Full pre-training (not fine-tuning or distillation from a GPU-trained teacher)
• Framing: Explicitly described as a proof-of-concept targeting hardware-agnostic AI development

How it works

At a conceptual level, the challenge VIDRAFT is solving is significant. Standard neural network training relies on backpropagation: computing gradients through the entire network and using those gradients to update weights. This process is computationally intensive, requires high numerical precision (typically FP32 or BF16), and assumes a memory architecture that supports both the forward pass and the backward pass simultaneously.

Inference-only NPUs are generally optimized for the forward pass only. They often rely on quantized arithmetic, fixed-function datapaths, and tightly constrained memory hierarchies — all of which create friction when you attempt to run a standard training loop.

What VIDRAFT appears to have done, at a high level, is adapt the training pipeline to operate within those constraints:

• Gradient computation had to be handled within the precision and datapath limits of the NPU's native operations
• Memory management during the backward pass would need to be restructured relative to the assumptions baked into typical GPU training frameworks
• Software abstraction is doing heavy lifting here — the hardware-level limitations are being worked around at the framework or compiler layer, not resolved at the silicon level

The source stops short of detailing the specific techniques used, and no internal hyperparameters, loss curves, or compiler internals are disclosed. The framing is proof-of-concept: demonstrating that the hardware barrier is a software constraint as much as a physical one.

Benchmarks & results

The source does not publish quantitative benchmark comparisons — no perplexity scores, downstream task accuracy numbers, or training throughput figures (tokens/second) are provided for VKORE-46M-NPU.

Qualitatively, the claim is that pre-training on inference-only NPUs reached convergence — meaning the model trained successfully and was not abandoned due to hardware failure or numerical instability. That convergence on this class of hardware is itself the result being reported.

What remains publicly unaddressed (and is worth watching for in follow-up reporting):

• Whether the loss landscape is comparable to a GPU-trained baseline of equivalent size
• Training wall-clock time relative to cloud GPU alternatives
• Energy consumption and thermal behavior over extended training runs
• Scalability trajectory toward billion-parameter models

How to try it

Public access to VKORE-46M-NPU has not been announced in this source. No Hugging Face repository, GitHub link, OpenAI-compatible API endpoint, or model weights download is mentioned in the coverage.

If VIDRAFT follows its established pattern with other VKORE-series models, a Hugging Face release may follow the proof-of-concept announcement — but that is not confirmed here. Watch the VIDRAFT Hugging Face organization page and their official channels for updates.

FAQ

Q: Does "inference-only NPU" mean the chip literally cannot do training, or just that it wasn't designed for it?
A: The distinction matters. "Inference-only" refers to the chip's intended design purpose and typical optimization targets — not necessarily a hard physical impossibility. VIDRAFT's work suggests the hardware can be coaxed into running training workloads through software-level adaptation, even if it was never architected with backpropagation in mind.

Q: Is a 46M-parameter model large enough to validate the approach for production use cases?
A: At 46 million parameters, VKORE-46M-NPU is orders of magnitude smaller than the multi-billion-parameter models that define current state-of-the-art deployments. The honest answer is: not yet, at this scale. VIDRAFT frames this explicitly as a proof-of-concept. The critical open question — which the source does not answer — is whether the approach scales to 1B+ parameters without prohibitive training time or quality degradation compared to GPU baselines.

---

Originally reported by Mukta AI(글로벌) (2026-10-14) — source article.