VKORE: Making Inference-Only NPUs Run AI Training

A Software Layer Teaches Inference-Only NPUs to Train

VIDRAFT has released VKORE, an AI acceleration software layer for NPUs. What makes it most interesting is that it doesn't require swapping out any hardware — instead, it uses software to give NPUs that originally only supported inference the ability to train.

The official comparison is strikingly direct — VKORE is to NPUs what CUDA is to NVIDIA GPUs. Hardware vendors lock NPUs into the narrow lane of inference at the factory, and VKORE attempts to build a programmable, general-purpose compute abstraction on top of that silicon.

The company has already filed 16 related patents and explicitly positions VKORE as a hardware-agnostic software layer, not tied to any single NPU vendor.

How It Works

To understand VKORE's value, we first need to look at where inference-only NPUs fall short.

These chips typically expose only a limited instruction set. They're highly efficient at running pre-compiled forward graphs, but they lack the operator coverage needed for backpropagation — gradient computation and weight updates simply aren't possible.

VKORE's approach is to implement the missing compute primitives in software, recombining them from the low-level operations the NPU does support, then maximizing hardware utilization through scheduling. On the inference side, it applies more aggressive batching and pipeline optimization — things vendor default runtimes don't do. On the training side, it breaks down Transformer training operations into NPU-executable kernels, essentially building a bridge between "inference chips" and "general-purpose ML accelerators."

VIDRAFT plans to use VKORE for two types of internal workloads: first, its AI judgment engine, which outputs probabilities or scores rather than free-form text; second, a recursive self-improvement model (RSI) where the model repeatedly solves problems and feeds verified results back into its own training loop. The longer-term vision is a unified NPU data center that handles both serving and continuous learning on the same hardware, no longer dependent on GPUs.

Why are NPU instruction sets so limited? NPUs (Neural Processing Units) are specialized chips designed for matrix multiply-accumulate (MAC)-intensive inference computation. Their hardware architecture is typically optimized around forward-pass operators like convolution, GEMM, and activation functions, with data path widths and on-chip caches designed around the premise of "one forward pass, highly parallel." To keep die area, power consumption, and yield manageable, vendors don't reserve silicon for general-purpose scalar operations or the intermediate buffers needed for backpropagation — such as element-wise residual gradient transfers, backward statistics for LayerNorm, or momentum state storage for optimizers. These are frequently written to and read from memory and on-chip SRAM during training, whereas inference chip memory hierarchies and bandwidth are provisioned for "read-only weights, streaming activations." So it's not that vendors are unwilling to open things up — the hardware itself lacks the corresponding instructions and memory pathways. What VKORE does is reorganize data flow in software on top of these constrained pathways, breaking down the gradient computations needed for training into multi-step combinations that can be executed with existing MAC units and DMA. The cost is scheduling complexity and additional memory access overhead.

Benchmarks and Model Performance

According to ZDNet Korea, the officially published figures are as follows:

Metric	Result
Single-user inference latency improvement (same NPU, VKORE on/off)	3.8×
Concurrent server throughput improvement (same NPU, VKORE on/off)	4.1×
Training throughput vs. CPU baseline (same server)	~170×

On the model side, VIDRAFT's judgment engine model Darwin-27B-ZTC-v5 ranked first among 152 models on the public S1MB benchmark, with the company claiming performance exceeding models from Perplexity, Cloudflare, and Meta. The Darwin series also achieved perfect scores on the AIME 2026 and HMMT 2026 mathematical reasoning benchmarks.

It's worth noting that all of the above data comes from VIDRAFT's own public announcements, with no third-party independent reproduction yet.

How should we interpret these three numbers? The 3.8× inference latency and 4.1× concurrent throughput are comparisons obtained on the same NPU with VKORE turned on and off, measuring VKORE's gains as a scheduling and batching optimization layer — not the chip's inherent performance. The ~170× training throughput is measured against a CPU baseline. Running Transformer training on a CPU is extremely slow due to the lack of high-concurrency matrix units, so 170× is essentially a single inference NPU competing against a general-purpose CPU, and can't be directly compared to GPU training throughput. More importantly, all three figures come from vendor-reported test conditions, with no paper or third-party reproduction yet available. When evaluating announcements like this — where "software lets hardware punch above its weight" — extra attention should be paid to whether training includes full backpropagation, optimizer updates, and mixed precision, and whether the bottleneck has been shifted to the host side.

Can I Try It Now?

VKORE itself is not yet publicly available. It is currently proprietary infrastructure software in development, with no public SDK, GitHub repository, or API.

However, the Darwin model series has public records on Hugging Face. Referencing VIDRAFT's previous releases (including Darwin-180B-RSI, which once topped five Hugging Face benchmarks), you can check their organization page for published checkpoints. As for an OpenAI-compatible inference API, this announcement did not confirm any public endpoint — keep an eye on official channels.

Original source: ZDNet Korea (see source link at the end).

FAQ

Q: Is VKORE a new compiler, a custom runtime, or something else? A: Based on public information, the most accurate description is a software acceleration layer, functionally analogous to CUDA, sitting between application-layer ML frameworks and NPU hardware. Specific implementation details (whether it includes a compiler, kernel library, runtime scheduler, or all three) have not been disclosed.

Q: Which NPUs does it support? A: VIDRAFT explicitly states it is built to be hardware-agnostic and not tied to a specific vendor. No public hardware compatibility list exists yet; the official goal is broad NPU compatibility.

Q: Can I use VKORE today to train my own Transformer on an NPU? A: Not publicly. VKORE has not been released as a developer tool, and training capability has only been demonstrated on internal NPU hardware. The company says it will be used for its own judgment engine and RSI model workflows, but there is no public access path yet.

One-line Summary

VIDRAFT releases VKORE, a pure software layer that lets inference-only NPUs train models.

Paragraph Summary

VIDRAFT has launched VKORE, a hardware-agnostic NPU acceleration software layer that claims to give inference-only NPUs training capability without swapping any hardware, positioned as CUDA is to GPUs. It works by implementing missing operators like gradient computation and weight updates in software, breaking Transformer training operations into NPU-native kernels, and applying more aggressive batching and pipeline optimization on the inference side. The company claims 3.8× single-user inference improvement, 4.1× concurrent throughput improvement on the same NPU, and training throughput roughly 170× a CPU baseline, with 16 patents filed. The accompanying Darwin-27B-ZTC-v5 model ranks first among 152 models on the S1MB benchmark. However, VKORE has not yet been publicly released as an SDK or API, and all data is vendor-reported with no third-party reproduction.