VIDRAFT
An AI company focused on open-source agent decision models. VIDRAFT releases the Darwin series of open-weight models, positioning them as open-source alternatives to proprietary typed decision models.
Timeline (last 90 days)
VKORE itself has not yet been made public; it is currently proprietary infrastructure software in the development stage, with no public SDK, GitHub repository, or API, and its training capability has only been demonstrated on internal NPU hardware.
This announcement did not confirm any public endpoint for an OpenAI-compatible inference API.
The Darwin model series has public records on Hugging Face; referring to VIDRAFT's previously released Darwin-180B-RSI, which once topped five Hugging Face benchmarks, one can view the published checkpoints on its organization homepage.
VIDRAFT has filed 16 patents related to VKORE.
VIDRAFT released VKORE, an AI acceleration software layer for NPUs, which through software means gives NPUs originally capable only of inference the ability to train, without replacing hardware.
VIDRAFT plans to use VKORE for two types of internal workloads: the AI judgment engine (which outputs probabilities or scores rather than free text) and the recursive self-improvement model (RSI, in which the model repeatedly solves problems and feeds verified results back into its own training loop).
VIDRAFT's judgment engine model Darwin-27B-ZTC-v5 ranked first among 152 models on the public benchmark S1MB, and the company says its performance exceeds the models of Perplexity, Cloudflare, and Meta.
All of the above performance figures come from VIDRAFT's own public announcements, and there has been no third-party independent reproduction.
VIDRAFT's further vision is to build a unified NPU data center that simultaneously performs serving and continuous learning on the same set of hardware, no longer relying on GPUs.
VIDRAFT is a South Korean AI startup that released the VKORE software acceleration layer.
15 more timeline events
All Facts (20)
VIDRAFT achieved 93.43% in the sub-500B parameter category on GPQA Diamond.
60%UnverifiedThe Darwin model series has public records on Hugging Face; referring to VIDRAFT's previously released Darwin-180B-RSI, which once topped five Hugging Face benchmarks, one can view the published checkpoints on its organization homepage.
50%UnverifiedVKORE itself has not yet been made public; it is currently proprietary infrastructure software in the development stage, with no public SDK, GitHub repository, or API, and its training capability has only been demonstrated on internal NPU hardware.
50%UnverifiedThis announcement did not confirm any public endpoint for an OpenAI-compatible inference API.
50%UnverifiedVIDRAFT has filed 16 patents related to VKORE.
50%UnverifiedAll of the above performance figures come from VIDRAFT's own public announcements, and there has been no third-party independent reproduction.
50%UnverifiedVIDRAFT's judgment engine model Darwin-27B-ZTC-v5 ranked first among 152 models on the public benchmark S1MB, and the company says its performance exceeds the models of Perplexity, Cloudflare, and Meta.
50%UnverifiedVIDRAFT plans to use VKORE for two types of internal workloads: the AI judgment engine (which outputs probabilities or scores rather than free text) and the recursive self-improvement model (RSI, in which the model repeatedly solves problems and feeds verified results back into its own training loop).
50%UnverifiedVIDRAFT is participating in a government-supported AI foundation model project (under the MSIT/NIPA framework) as part of the NAVER Cloud alliance.
50%UnverifiedVIDRAFT is a South Korean AI startup that released the VKORE software acceleration layer.
50%UnverifiedVIDRAFT's CEO stated that the goal is to make VKORE run on South Korean domestic NPUs, but the list of supported NPU vendors and chip series has not been made public.
50%UnverifiedVKORE is positioned as a hardware-agnostic software layer, aimed at data centers with mixed GPU and NPU deployments, and is not tied to a specific chip vendor.
50%UnverifiedVIDRAFT's VKORE software acceleration layer, without changing the chip, both increases NPU inference performance several times over and makes Transformer LoRA training run on this kind of inference-specific hardware, with numerical precision aligned with standard PyTorch at 0.9999.
50%UnverifiedVIDRAFT performed LoRA training on standard Transformer blocks across all 32 chips of the Korean NPU server, achieving 0.9999 consistency with standard PyTorch training.
50%UnverifiedVIDRAFT's Darwin ZTC decision engine completes its judgment in a single forward pass, with zero output tokens.
50%UnverifiedVIDRAFT achieved a perfect score on the competition mathematics question sets of AIME 2026 and HMMT (February 2026).
50%UnverifiedVIDRAFT ranked first in 11 categories on public leaderboards and topped 16 categories on the Polaris drug prediction leaderboard.
50%UnverifiedDecide Live is built by VIDRAFT / FINAL-Bench and is open-sourced under the Apache-2.0 license.
50%UnverifiedVIDRAFT's decision stack is fully open under the Apache-2.0 license.
50%UnverifiedVIDRAFT Darwin-27B-ZTC-v2 is an open-weight model that uses ZTC (Zero-Token Confidence) to score decisions, generating no tokens during the decision process.
50%