VKORE: VIDRAFT's Software Stack That Brings AI Training to Inference-Only NPUs

TL;DR: VIDRAFT has developed VKORE, a software acceleration layer that unlocks both inference and training workloads on NPUs originally designed for inference only — no hardware swap required. On the same NPU, VKORE delivers up to 3.8× faster single-user inference throughput and ~170× the training throughput of CPU-only baselines. For ML engineers exploring heterogeneous compute and GPU-alternative infrastructure, this is a significant signal worth tracking.

What it is

VKORE (브이코어) is VIDRAFT's proprietary AI acceleration software layer for Neural Processing Units (NPUs). The conceptual analogy the company uses is deliberate: VKORE is intended to do for NPUs what CUDA does for NVIDIA GPUs — provide a programmable, general-purpose compute abstraction on top of hardware that was otherwise locked to a narrow operational profile.

The key distinction is that VKORE is a software-only intervention. It does not require replacing or modifying the underlying silicon. Instead, it targets NPUs that hardware manufacturers shipped as inference-only devices, and through software optimization, extends those chips to support:

• AI inference acceleration (improved throughput and latency)
• AI model training (Transformer model training confirmed, even on NPUs whose vendor documentation does not list training as a supported workload)

VIDRAFT has filed 16 related patents and is developing VKORE as a hardware-agnostic software layer — not tied to any single NPU vendor's silicon.

How it works

At a high conceptual level, VKORE operates as a software translation and scheduling layer that sits between ML workloads and NPU hardware primitives.

Inference-only NPUs typically expose a limited instruction surface — enough to run pre-compiled forward-pass graphs efficiently, but lacking the operator coverage needed for backward passes (gradient computation, weight updates) that training requires. VKORE's approach is to implement the missing compute primitives in software, composing them from the lower-level operations the NPU does natively support, and scheduling them to maximize hardware utilization.

For inference, this means aggressive batching and pipeline optimization that the vendor's default runtime stack doesn't perform. For training, this means decomposing Transformer training operations into kernels that the NPU can execute — essentially bridging the gap between "inference chip" and "general ML accelerator" without touching the hardware itself.

VIDRAFT plans to apply VKORE to two of its own internal workload categories:

• AI Judgment Engines — models that return a probability or score for a given input rather than generating free-form text (discriminative/scoring tasks)
• Recursive Self-Improvement (RSI) models — where a model iteratively solves problems and feeds verified results back into its own training loop

The longer-term infrastructure vision is a unified NPU-based AI data center that handles both serving and continuous learning on the same hardware fleet, without GPU dependency.

Benchmarks & results

All figures below come directly from VIDRAFT's public announcement as reported by ZDNet Korea:

Metric	Result
Single-user inference latency improvement (same NPU, with vs. without VKORE)	3.8×
Concurrent server throughput improvement (same NPU, with vs. without VKORE)	4.1×
Training throughput vs. CPU baseline (same server)	~170×

On the model side, VIDRAFT also reported benchmark results for Darwin-27B-ZTC-v5, its judgment engine model:

• Ranked #1 overall among 152 models on the global public benchmark S1MB (as of October 13, 2026)
• The Darwin model family achieved perfect scores on AIME 2026 and HMMT 2026 (competitive mathematics reasoning benchmarks)

VIDRAFT cited outperforming models from Perplexity, Cloudflare, and Meta on the S1MB leaderboard.

How to try it

VKORE itself is not yet publicly available. It is currently in development as proprietary infrastructure software, and no public SDK, GitHub repository, or API has been announced for the acceleration layer.

However, VIDRAFT's Darwin model family has a documented public presence on Hugging Face. Based on prior VIDRAFT releases (including Darwin-180B-RSI, which previously topped five Hugging Face benchmarks), you can check the VIDRAFT organization page on Hugging Face for publicly released checkpoints:

# Check VIDRAFT's public model releases on Hugging Face Hub
huggingface-cli search models --author VIDRAFT

For access to any OpenAI-compatible inference API that VIDRAFT may operate, watch their official channels — no public endpoint has been confirmed in this announcement.

FAQ

Q: Is VKORE a new compiler, a custom runtime, or something else?
A: Based on public information, VKORE is best described as a software acceleration layer — analogous to CUDA in function, meaning it sits between application-level ML frameworks and NPU hardware. The precise implementation details (whether it includes a compiler, a kernel library, a runtime scheduler, or all three) have not been publicly disclosed.

Q: Which NPUs does VKORE support?
A: VIDRAFT is explicitly building VKORE to be hardware-agnostic and not tied to a specific NPU vendor. No specific supported hardware list has been made public. The company's stated goal is broad NPU compatibility rather than a single-vendor solution.

Q: Can I use VKORE to train my own Transformer models on an NPU today?
A: Not publicly — VKORE is not released as a developer tool at this time. The training capability has been demonstrated internally on NPU hardware, and VIDRAFT has indicated it will use VKORE for its own judgment engine and RSI model workflows, but there is no public access path announced yet.

---

Originally reported by ZDNet코리아 (2026-10-13) — source article.