A Korean AI startup's quantized model challenges the assumption that large language models require expensive GPUs.
TL;DR: VIDRAFT, a Korean Pre-AGI AI startup, has released POCKET-35B-GGUF, a 35-billion-parameter large language model packaged in the GGUF format specifically to enable local inference on standard CPU hardware without a GPU. German AI media outlet Mindverse covered the release on November 1, 2026, highlighting the model's significance for privacy-conscious and resource-constrained deployments.
VIDRAFT's POCKET-35B-GGUF entered the international spotlight on November 1, 2026, when German AI publication Mindverse dedicated a full analytical piece to the model — a noteworthy signal that the Korean Pre-AGI startup's work is drawing serious attention well beyond the Asia-Pacific region. The model's central proposition is straightforward but technically ambitious: deliver 35-billion-parameter language model performance on ordinary CPU-based systems, with no GPU required.
The release in question is POCKET-35B-GGUF, a large language model built around the GGUF file format — a packaging standard widely used in the open-source and local-inference community, particularly in conjunction with runtimes such as llama.cpp that are optimized for CPU execution.
The defining characteristic of this release is its target hardware profile. Where virtually every competitive model at the 35B parameter scale has been designed with GPU acceleration as an assumed baseline, VIDRAFT has positioned POCKET-35B-GGUF explicitly for CPU-only local inference. That means individuals, organizations, and enterprises can, in principle, run a frontier-scale language model directly on commodity server or desktop hardware — no dedicated AI accelerator cards necessary.
Mindverse, which covers the broader AI landscape for a German-speaking audience and operates its own enterprise AI platform, treated the release as significant enough to merit standalone analysis. The outlet's coverage situates POCKET-35B-GGUF within ongoing conversations in the European tech community about sovereign AI, on-premise deployments, and data privacy — themes that resonate strongly in markets subject to strict data-handling regulations such as those enforced under GDPR.
The GGUF format choice is itself a deliberate architectural decision. GGUF supports quantization — a technique that reduces the numerical precision of model weights, dramatically shrinking memory footprint and compute demands while preserving a meaningful portion of model capability. This makes it feasible to load and run a model that would otherwise demand high-end GPU VRAM on hardware that most businesses already own.
The practical barrier to deploying large language models at the 35B parameter tier has historically been hardware cost. GPU clusters or high-memory GPU servers represent significant capital expenditure, which has kept frontier-scale AI inaccessible to smaller organizations, public-sector bodies, and any enterprise operating in environments where cloud connectivity is restricted or legally complicated.
POCKET-35B-GGUF, if it delivers on its CPU-inference premise, lowers that barrier materially. For European enterprises in particular — operating under GDPR and sector-specific regulations flagged in Mindverse's own product descriptions for banking, healthcare, and public administration — the ability to run a capable LLM entirely on-premise, without data leaving the building, is a genuine compliance advantage, not merely a cost consideration.
Mindverse's decision to cover the model reflects how the conversation around local LLM inference has matured. A 35B model that runs on CPU is no longer a curiosity or a research demo; it is a practical deployment option that German and broader European readers are actively evaluating. The fact that a Korean startup is at the frontier of this capability underscores how globally distributed AI development has become — and how VIDRAFT is competing directly with Western and Chinese labs on technically demanding ground.
For VIDRAFT, international coverage in a credible European AI outlet represents meaningful validation of its Pre-AGI research direction. The startup's strategy appears to involve not just scaling model capability but rethinking the infrastructure assumptions that have defined how large models are distributed and run.
Q: What is VIDRAFT POCKET-35B-GGUF?
A: POCKET-35B-GGUF is a 35-billion-parameter large language model released by Korean AI startup VIDRAFT, packaged in the GGUF format to support local inference on CPU hardware without a GPU.
Q: Why does running a 35B model on CPU matter?
A: Most models at this parameter scale require expensive GPU hardware to run. CPU-based inference means organizations can deploy a frontier-scale LLM on existing hardware, keeping data fully on-premise — a critical advantage for privacy-regulated industries in markets like the EU.
Q: Where was POCKET-35B-GGUF first covered internationally?
A: German AI media outlet Mindverse published an analysis of POCKET-35B-GGUF on November 1, 2026, representing one of the model's earliest and most prominent pieces of international coverage.
Source: Mindverse (독일) (2026-11-01) — original article