Korean AI startup VIDRAFT proves frontier-scale models no longer require data-center hardware.
TL;DR: VIDRAFT, a Korean AI deeptech startup, has publicly released a compressed version of its 180-billion-parameter Darwin model that can run on a consumer gaming laptop with 8 GB of GPU memory and 32 GB of RAM. The lightweight release, called POCKET-Darwin-180B-GGUF, reduces the original 360 GB model file to 111 GB through 4-bit quantization while preserving a Mixture-of-Experts architecture. VIDRAFT simultaneously published the model on Hugging Face and China's ModelScope on October 6, 2026.
VIDRAFT, the Korean Pre-AGI AI startup led by CEO Kim Min-sik, made headlines on October 6, 2026, when it unveiled a compressed, consumer-ready edition of its flagship 180-billion-parameter Darwin model — one capable of running on hardware that costs roughly 250 times less than the server infrastructure the original model typically demands.
The base model behind the release is Darwin-180B-RSI. At BF16 precision, its model files alone weigh in at 360 GB, and running it conventionally requires a multi-GPU server configuration that puts it firmly in data-center territory.
VIDRAFT's engineering team tackled that barrier head-on. The result is POCKET-Darwin-180B-GGUF, a 111 GB package distributed in the GGUF file format — the same format used by the open-source llama.cpp inference engine. Users can download four files totalling 111 GB and launch the model with a single command in a current version of llama.cpp. No cloud account and no dedicated GPU server are required.
The compression strategy combines two techniques. First, the team applied 4-bit quantization to the model weights, dramatically shrinking the file size without rebuilding the underlying architecture. Second, and equally important, the model retains the original Darwin-180B's Mixture-of-Experts (MoE) structure: out of 512 specialized sub-networks, only 10 are activated per token during inference. That means roughly 3 billion parameters do the actual compute work at any given moment, even though the full 180-billion-parameter structure remains intact. On top of that, VIDRAFT integrated an SSD-streaming approach based on llama.cpp so that only the expert weights needed for a given inference step are loaded from storage — the entire model never has to sit in memory at once.
The practical hardware floor is notably low. A gaming laptop priced in the roughly two-million-Korean-won range — equipped with 8 GB of GPU memory and 32 GB of RAM — is sufficient to run the model, achieving approximately 4.17 tokens per second. On a mini-PC with 128 GB of RAM, the full model can be loaded directly into system memory and run without any GPU at all. When VIDRAFT tested a single 16-thread server CPU without a GPU, generation speed came in at 18.4 to 21.0 tokens per second, with memory usage of 78.8 GB.
On the benchmark front, VIDRAFT compared performance on 2,000 questions from the MMLU-Pro comprehensive knowledge test before and after compression. Both the original and the quantized model recorded the same score, suggesting that 4-bit quantization did not meaningfully degrade accuracy for this evaluation.
POCKET-Darwin-180B-GGUF was published simultaneously on Hugging Face and China's ModelScope.
Compressing large language models is not new — 4-bit quantization and MoE architectures have each been explored by multiple organizations. What makes VIDRAFT's release notable, according to the AI타임스 report, is that publicly available cases of a 180-billion-parameter-class model maintaining its full architecture while being made runnable on consumer PCs and CPUs are rare in South Korea and uncommon globally.
The practical implications are significant for privacy-sensitive deployments. Because the model runs entirely on local hardware, organizations that cannot or will not send data to external cloud services — including enterprises handling confidential information and public-sector institutions — gain a credible on-premise option at a fraction of traditional infrastructure cost. VIDRAFT specifically cites these segments as primary target users.
CEO Kim Min-sik framed the dual strategy concisely, stating that boosting performance through Recursive Self-Improvement (RSI) and then making the resulting model runnable anywhere through the POCKET line is the company's core competitive proposition.
Q: What is POCKET-Darwin-180B-GGUF and who made it?
A: POCKET-Darwin-180B-GGUF is a 4-bit quantized, 111 GB version of VIDRAFT's 180-billion-parameter Darwin-180B-RSI model. It was developed and released by Korean AI startup VIDRAFT on October 6, 2026.
Q: What hardware do you need to run VIDRAFT's 180B model locally?
A: At minimum, a gaming laptop with 8 GB of GPU memory and 32 GB of RAM can run the model. A mini-PC with 128 GB of RAM can load and run it entirely in system memory without a GPU.
Q: Does compressing the Darwin-180B model reduce its accuracy?
A: Based on VIDRAFT's own evaluation using 2,000 MMLU-Pro questions, the accuracy score was the same before and after 4-bit quantization was applied.
Source: AI타임스 (2026-10-06) — original article