FINAL-Bench/POCKET-26B-GGUF

🤗 Hugging Face sourcetext-generationapache-2.028 GBGGUF✓ 2 checksumsupdated today
Submit in one command

Run it next to your model folder. It makes the torrent, checks your files against Hugging Face, and submits it. You just start seeding and paste your key from your account. It only reads your files and never changes them. Read the script first if you like.

curl -fsSL https://pirateface.co/package.sh | bash -s -- --repo FINAL-Bench/POCKET-26B-GGUF ./model-folder
Needs a seeder →

🆕 POCKET-Qwen3.8-Flash-Next — a 180B model running on a laptop with 8 GB VRAM + 32 GB RAM · 4.17 tok/s measured.

🆕 POCKET-Zimage-CPU — photoreal images in 46 s on a CPU only. No GPU, no CUDA, no Python.

📚 Collections

▶ POCKET Models — this family (on-device, no GPU) Darwin Family · Aether Foundation · VKAE Accelerated

POCKET-26B-GGUF · 한국어

A Gemma4-26B-A4B-based pocket model that loads in any app today — Ollama, LM Studio, PocketPal — with no bleeding-edge runtime needed. Korean-tuned, GPU-optional.

🚀 Try it live on a CPU (no GPU), no install →

Pick your build →

Why this one?

POCKET-26B takes Google's Gemma4-26B-A4B (25.2B total, ~4B active MoE, Apache-2.0) and re-quantizes it with our proprietary Korean-tuned quantization — unpruned, so quality holds. Unlike our Qwen-based POCKET (which needs a very recent llama.cpp build for its qwen35moe architecture), Gemma4 loads in every mainstream runtime today: Ollama, LM Studio, PocketPal, koboldcpp, and the browser.

Quality — GPQA-Diamond, greedy, 198 questions (our harness)

Build GPQA-Diamond vs base
Gemma4-26B-A4B (base) 67.7% —
POCKET-26B Q4_K_M 67.7% = base (lossless)
POCKET-26B Q2_K (mixed) ⭐ 67.2% −0.5pp (≈ lossless)

Single greedy pass, 198 items → ±~3 pp noise. Our proprietary Korean-tuned quantization is statistically lossless vs the base.

Files in this repo

File Size Runs on Best for
POCKET-26B-Q4_K_M.gguf 17 GB PC / high-RAM top quality
POCKET-26B-Q2_K.gguf ⭐ 11 GB 12 GB phone / PC / browser universal daily driver

Our mixed-precision quantization keeps the most quality-critical weights at higher precision — that is why Q2_K holds 67.2% while a plain uniform Q2 collapses to ~44%.

Quickstart — loads anywhere

# stock llama.cpp — brew / winget / apt, or LM Studio / Ollama / PocketPal
llama-cli -m POCKET-26B-Q2_K.gguf -p "대한민국의 수도는?" -ngl 0 -t 8

No fork, no bleeding-edge build — Gemma4 support has shipped in every mainstream runtime since April 2026.

Lineage (honest)

Based on google/gemma-4-26B-A4B-it (Apache-2.0). We do not re-host it unchanged — we add our proprietary Korean-tuned quantization (VIDRAFT). We deliberately do not prune it: Gemma4's low-bit robustness collapses under pruning (measured), so we keep all 128 experts and win on quality + universal compatibility instead.

Limitations

  • For 8 GB phones (~5 GB budget), use POCKET-KR-GGUF (5.1 GB) — Gemma4 cannot be shrunk that far without collapse.
  • On-device iPhone/Mac throughput not yet measured by us — community reports welcome.

Learn more

License

Apache-2.0 — use, modify, redistribute freely.


POCKET is a VIDRAFT model family. Runs anywhere, no GPU.


🧩 The POCKET Family — On-device AI by VIDRAFT

Big models, small hardware. No GPU, no cloud.

Models

Demos & tools (Spaces)

📚 Full POCKET collection