DEV Community

AI OpenFree
AI OpenFree

Posted on

VIDRAFT POCKET-35B: A 35B-Parameter On-Device LLM That Hits 1M Hugging Face Downloads in 7 Weeks — No GPU Required

#ai #opensource #llm #machinelearning

VIDRAFT POCKET-35B: A 35B-Parameter On-Device LLM That Hits 1M Hugging Face Downloads in 7 Weeks — No GPU Required

TL;DR: VIDRAFT's POCKET-35B is a 35-billion-parameter language model designed to run entirely on consumer hardware — laptops, mini-PCs, and tablets — without a dedicated GPU. It reached 1 million cumulative Hugging Face downloads in roughly seven weeks after release, making it the fastest Korean AI model to hit that milestone. If you're building on-device or offline-capable AI pipelines, this model is worth evaluating.

What it is

POCKET-35B is an on-device large language model (LLM) developed by VIDRAFT, a Korean AI foundry startup. Key facts from the source:

  • Parameter count: 35 billion
  • Release date: July 22, 2026
  • Target hardware: Consumer-grade devices — laptops, mini-PCs, tablets — with no discrete GPU required
  • Deployment model: Fully local / on-device inference; no external server connection needed
  • Primary distribution format: GGUF (the quantized format used by runtimes like llama.cpp for CPU-based inference)
  • Developer: VIDRAFT, which describes itself as an AI foundry company

The key design constraint is explicit: POCKET-35B is built to run in environments where cloud calls aren't possible or desirable — offline-first, privacy-sensitive, or bandwidth-constrained deployments.

How it works

At a conceptual level, running a 35B-parameter model on CPU-only hardware is made practical through post-training quantization, specifically the GGUF format. GGUF (originally from the llama.cpp ecosystem) stores model weights in reduced-precision integer formats (e.g., 4-bit or 8-bit), dramatically shrinking memory footprint and enabling inference on unified-memory consumer hardware without a GPU.

The on-device architecture means:

  • All computation happens locally. Tokens are generated on-device, so there is no network round-trip latency and no data leaves the machine.
  • No server infrastructure is required at runtime. Developers ship the model file alongside their application, or users download it once.
  • CPU-based runtimes (such as llama.cpp and its wrappers) handle the quantized matrix multiplications across available CPU cores and, where available, integrated GPU units like Apple Silicon's Neural Engine or integrated Intel/AMD graphics.

VIDRAFT has not published specifics about the base architecture or training approach beyond what is in the public release — this article reflects only what has been disclosed.

Benchmarks & results

The source article provides download and adoption metrics rather than task-specific benchmark scores. Here is what was publicly reported:

  • 1,076,570 cumulative downloads on Hugging Face as of September 11, 2026 (approximately 7 weeks post-release)
  • ~19,936 average daily downloads over the full period since release
  • 76% of all downloads (815,008) occurred in the most recent 30-day window, indicating accelerating adoption — not a spike-and-fade pattern
  • Fastest Korean AI model to cross the 1M download threshold among 21 Korean models that have done so; the #2 Korean model averaged ~9,468 downloads/day — roughly 2.1× slower
  • Global GGUF-format ranking: #13 worldwide for 30-day downloads, among 65 GGUF models in the global top-300 list (47 Chinese, 12 American, 2 Korean — both from VIDRAFT)
  • Overall global ranking: #103 by 30-day downloads across all formats in VIDRAFT's Global LLM Download Leaderboard
  • VIDRAFT's two POCKET models together account for 60.2% of total downloads across all Korean models in the global top-300

Note: These are distribution/adoption metrics. Task-specific benchmarks (reasoning, coding, instruction-following) have not been published in this article. Check the Hugging Face model card for any evaluation results VIDRAFT has posted directly.

How to try it

POCKET-35B is publicly available on Hugging Face. Because it ships in GGUF format, the typical workflow uses llama.cpp or a compatible frontend:

1. Browse/download from Hugging Face

Search for VIDRAFT/POCKET-35B (or the exact model ID listed on VIDRAFT's Hugging Face profile) and download the GGUF quantization variant appropriate for your available RAM.

# Example using the Hugging Face CLI (install once)
pip install huggingface_hub

huggingface-cli download VIDRAFT/POCKET-35B --include "*.gguf"
Enter fullscreen mode Exit fullscreen mode

⚠️ Verify the exact repository name on huggingface.co/VIDRAFT before running — use the canonical ID shown there.

2. Run locally with llama.cpp or Ollama

Any runtime that supports GGUF files can load the model. No GPU is required; CPU-only inference is the intended path.

No public API endpoint or GitHub repository was mentioned in this source article. Check VIDRAFT's Hugging Face organization page and official channels for the latest access options.

FAQ

Q: How much RAM do I need to run POCKET-35B without a GPU?
A: This depends on the quantization level of the GGUF file you download. Lower-bit quantizations (e.g., Q4) require significantly less RAM than higher-bit variants. Check the individual GGUF file sizes on the Hugging Face repository and match them to your available system RAM — as a rule of thumb, the file size is roughly the minimum RAM needed.

Q: Does "no GPU required" mean GPU acceleration is unsupported, or just unnecessary?
A: It means the model is designed to function correctly on CPU alone. Most GGUF-compatible runtimes can also offload layers to an integrated or discrete GPU if one is available, which will improve throughput. "No GPU required" is a minimum-requirement statement, not a ceiling.

Q: Is POCKET-35B multilingual, and does it support Korean?
A: The source article does not specify language coverage. Given VIDRAFT is a Korean AI company and the model has high adoption in the Korean developer community, Korean language support is likely — but check the model card on Hugging Face for the official language list.

Q: Are there smaller POCKET models for more constrained hardware?
A: The article references "two POCKET models" appearing in global download rankings, suggesting at least one other variant exists. Check VIDRAFT's Hugging Face organization page for the full model lineup.

Originally reported by ZDNet Korea (2026-09-15) — source article.

Top comments (0)