Proving that frontier-scale AI doesn't require frontier-scale hardware
TL;DR: VIDRAFT's on-device model suite POCKET has surpassed 3.15 million total downloads on Hugging Face as of November 1, 2026. The flagship POCKET-Darwin-180B model runs on a laptop with just 8 GB of VRAM and 32 GB of system RAM, achieving 4.17 tokens per second. All models load in the official, unmodified version of llama.cpp with no cloud dependency required.
VIDRAFT, the Korean Pre-AGI AI startup, reached a significant milestone on November 1, 2026, when its POCKET family of on-device large language models crossed 3,154,913 cumulative downloads across three primary Hugging Face repositories — a figure reported by Chinese internet portal NetEase (网易) based on publicly available Hugging Face API data.
According to the NetEase report, POCKET's three main Hugging Face repositories logged a combined 1,113,164 downloads in the most recent 30-day window alone. POCKET-Darwin-180B, the largest model in the lineup, had been live for fewer than 30 days at the time of reporting, meaning its 30-day figure and its all-time total are identical.
POCKET is a collection of GGUF-format model files built around three core principles: standard toolchain compatibility, sparse Mixture-of-Experts (MoE) architecture, and benchmarking based on real hardware measurements rather than theoretical estimates.
Standard toolchain compatibility means every POCKET file loads directly into the official release of llama.cpp, as well as the popular front-ends built on top of it — Ollama, LM Studio, and PocketPal — without any modified or forked runtime. CUDA is not required, and none of the models phone home to a cloud service.
Sparse MoE is what makes the size-to-speed ratio possible. Because only a small fraction of network parameters are activated per token, models at the 35B and 180B scale can generate text at usable speeds on CPU-only machines.
Real-hardware benchmarks underpin every speed claim. On a 16-thread Xeon CPU running a standard llama.cpp build, POCKET-35B delivered 27.0 tokens per second, compared with 10.1 tok/s for a comparable 27B model in the same test environment. On an H100, those figures rose to 197 tok/s and 89 tok/s respectively — though the competing model had a faster prompt-processing rate on that hardware, at 1,816 tok/s versus POCKET's 753 tok/s.
Quality under aggressive quantization is a common concern with compact formats. POCKET-26B uses mixed-precision quantization to retain nearly all of the base model's benchmark score even at 2-bit compression. POCKET-Darwin-180B's 4-bit variant matched its BF16 counterpart in quality evaluation, scoring 87.65%.
For users who want to run the models locally, the workflow is straightforward: download the GGUF file from Hugging Face, then launch it with llama-cli. Setting -ngl 0 forces pure CPU inference. One practical note flagged in the source: the -t thread parameter should be set to the number of physical cores, not logical threads — using all available threads can paradoxically slow generation significantly. The same GGUF files work without modification in Ollama, LM Studio, and PocketPal.
Concrete hardware requirements reported in the article:
For users who prefer not to install anything locally, a browser-based option called POCKET Image Studio is included in the POCKET collection.
The ability to run a 180-billion-parameter model on consumer-grade hardware — without cloud access, proprietary runtimes, or a discrete GPU — represents a meaningful shift in who can access frontier-scale AI. Most models in the 100B-plus range have required either powerful cloud infrastructure or specialized workstations. VIDRAFT's POCKET challenges that assumption by combining sparse MoE efficiency with rigorous real-world testing, then distributing the results through open, standard formats that the broader open-source ecosystem already supports.
The download velocity is itself a signal: more than one million downloads in a single month, across a model family that includes a 180B variant released less than 30 days prior, suggests rapid organic adoption among developers and enthusiasts unwilling to depend on API-gated or cloud-hosted alternatives.
Q: What hardware do you need to run VIDRAFT POCKET-Darwin-180B?
A: According to the NetEase report, the 4-bit version (111 GB) can run on a laptop equipped with 8 GB of VRAM and 32 GB of system RAM at approximately 4.17 tokens per second. A mini-PC with 128 GB of RAM can run it without any GPU at all.
Q: Does VIDRAFT POCKET require a custom or modified version of llama.cpp?
A: No. All POCKET GGUF files are compatible with the official release of llama.cpp, as well as front-ends such as Ollama, LM Studio, and PocketPal, with no modifications required.
Q: How many times has VIDRAFT POCKET been downloaded?
A: Hugging Face public API data cited by NetEase shows POCKET reached 3,154,913 total downloads across its three main repositories, with 1,113,164 of those occurring in the 30 days leading up to November 1, 2026.
Source: 넷이즈 网易 (중국) (2026-11-01) — original article