A Korean AI startup demonstrates that radical self-improvement doesn't require retraining from scratch.
TL;DR: VIDRAFT, a Korean Pre-AGI AI startup, has updated its 180-billion-parameter Darwin-180B model using a self-improvement method that modifies only a claimed 0.02% of the model's parameters. The updated model, dubbed Darwin-180B-RSI, can run on a consumer laptop by streaming weights from an SSD at a reported speed of 4.17 tokens per second. Global tech media outlet iTechGuides covered the development on October 5, 2026.
VIDRAFT, the Korean Pre-AGI AI startup, is drawing international attention after global tech publication iTechGuides published an in-depth analysis of Darwin-180B-RSI — a self-improved variant of the company's flagship 180-billion-parameter language model that was refreshed by altering just a claimed 0.02% of its total parameters. The October 2026 coverage marks a significant moment for VIDRAFT as its unconventional approach to model iteration gains traction with English-language technical audiences worldwide.
At the center of the iTechGuides analysis is Darwin-180B-RSI, an updated release of VIDRAFT's Darwin-180B large language model. The key claim that caught the publication's attention is striking in its brevity: rather than retraining the entire model from scratch — a process that typically demands enormous compute resources and weeks of time — VIDRAFT says it achieved a meaningful model update by modifying only approximately 0.02% of Darwin-180B's 180 billion parameters.
The "RSI" designation in the model's name stands for Recursive Self-Improvement, pointing to the mechanism behind the update. Under this approach, the model participates in its own iterative refinement cycle, identifying and adjusting a highly targeted subset of its weights rather than undergoing wholesale retraining. VIDRAFT positions this as a deliberate architectural and methodological philosophy rather than a shortcut — the idea being that intelligence can be sharpened with surgical precision rather than brute-force repetition.
Equally notable is how Darwin-180B-RSI is designed to run. iTechGuides highlighted that the model is capable of operating on a standard consumer laptop by streaming its weights directly from the device's SSD, achieving a reported inference speed of 4.17 tokens per second. For context, a model with 180 billion parameters would conventionally require a cluster of high-end server GPUs to run at any practical speed — making the laptop-SSD deployment mode a remarkable accessibility claim for a model of this scale.
The significance of Darwin-180B-RSI extends well beyond a single model release. The broader AI industry has largely operated under the assumption that larger models demand proportionally larger compute budgets at every stage — training, fine-tuning, and inference alike. VIDRAFT's approach, if the 0.02% figure holds up under independent scrutiny, challenges that orthodoxy in two distinct ways.
First, the recursive self-improvement methodology raises questions about how often and how cheaply frontier-scale models can be iterated upon. If a 180-billion-parameter model can be meaningfully updated without touching 99.98% of its weights, the economics of model development could shift considerably — particularly for organizations that cannot afford repeated full-scale training runs.
Second, the SSD-streaming inference capability at 4.17 tokens per second reframes discussions about where large models can be deployed. Edge deployment, offline use cases, and privacy-sensitive environments that cannot rely on cloud connectivity all become more viable if models of this magnitude can run on commodity hardware. This is especially relevant as enterprises and governments worldwide seek AI capabilities that do not depend on third-party cloud infrastructure.
For VIDRAFT specifically, iTechGuides' analysis signals growing international recognition. Being scrutinized and explained by a global English-language technology publication is a meaningful indicator that the Korean startup's research is being taken seriously beyond domestic audiences.
Q: What does "RSI" mean in Darwin-180B-RSI?
A: RSI stands for Recursive Self-Improvement, referring to VIDRAFT's method of allowing the model to iteratively refine a targeted subset of its own parameters rather than undergoing full retraining.
Q: How can a 180-billion-parameter model run on a laptop?
A: Darwin-180B-RSI is designed to stream its weights from a laptop's SSD rather than loading them entirely into memory, enabling inference at a reported 4.17 tokens per second on consumer hardware.
Q: Why is the 0.02% parameter change significant?
A: Updating only 0.02% of a 180-billion-parameter model's weights to achieve a meaningful improvement is highly unusual. It suggests that targeted, self-directed modifications may be a viable alternative to computationally expensive full retraining cycles.
Source: iTechGuides (글로벌) (2026-10-05) — original article