๐ POCKET-Qwen3.8-Flash-Next โ a 180B model running on a laptop with 8 GB VRAM + 32 GB RAM ยท 4.17 tok/s measured.
๐ POCKET-Zimage-CPU โ photoreal images in 46 s on a CPU only. No GPU, no CUDA, no Python.
๐ Collections
โถ POCKET Models โ this family (on-device, no GPU) Darwin Family ยท Aether Foundation ยท VKAE Accelerated
POCKET-KR-GGUF ยท ํ๊ตญ์ด
35B์์ ํ๊ตญ์ด ์ ๋ฌธ๊ฐ๋ง ๊ณจ๋ผ๋ธ ํฐ์ฉ ๋น๋. Android 8 GB+์์ GPU ์์ด ๋๋๋ค. iPhone์ POCKET-KR-MLX๋ฅผ ๋ฐ์ผ์ธ์.
๐ Try it live, no install โ โ both answering on a CPU-only box (no GPU). POCKET-26B is Gemma4-based.
Pick your build โ
The POCKET lineup โ pick by your device
| Repo | File | Size | Runs on | Best for | Korean PPL* |
|---|---|---|---|---|---|
| POCKET-35B-GGUF | Q4_K_M |
21 GB | PC / server (32 GB RAM) | top quality | 5.79 |
| POCKET-35B-GGUF | Q2_K โญ |
13 GB | mini-PC, no GPU | daily driver | 6.49 |
| POCKET-35B-GGUF | IQ1_M |
8.2 GB | 16 GB RAM box | smallest full model | 9.69 |
| POCKET-KR-GGUF | IQ2_M |
5.1 GB | Android 8 GB+ | ๐ฐ๐ท Korean phone | 7.95 |
| POCKET-KR-MLX | 2-bit | 5.1 GB | ๐ iPhone / iPad / Mac | ๐ฐ๐ท Korean, Apple-native | 7.95 |
| POCKET-EN-GGUF | iPhone-mix |
5.3 GB | ๐ iPhone (PocketPal) | ๐ English phone | โ |
| POCKET-EN-GGUF | PC-mix |
6.8 GB | PC / Android | ๐ English, best quality | โ |
| POCKET-Qwen3.8-Flash-Next-GGUF | Q4_K_M |
111 GiB | ๐ป laptop, 8 GB VRAM + 32 GB RAM | 180B on a laptop | 6.03โ |
*Wikipedia-Korean perplexity, lower is better. Q4_K_M = 5.79 baseline. English builds are tuned on English; see each repo.
โ Separate 80-chunk run (40,960 tokens) on a different model โ compare within a model, not across rows.
๐ Why MLX for Korean but GGUF for English on iPhone? Apple-native MLX only does uniform quantization. Korean survives it; English needs our proprietary quantization, which only GGUF supports โ so the English iPhone build ships as a GGUF you run with PocketPal. Honest, not lazy.
๐ POCKET-26B โ a Gemma4-26B-A4B-based sibling that loads in any app today (Ollama ยท LM Studio ยท PocketPal ยท MLX), no bleeding-edge runtime needed: GGUF (
Q2_K11 GB ยทQ4_K_M17 GB ยท GPQA-Diamond 67%). Universal compatibility for 12 GB phones, PC, and browser.
Benchmarks โ what is measured, what is not
We measure Bonsai on the same machine with the same stock llama.cpp, and we tell you where we lose.
[measured] Generation speed โ POCKET wins on both CPU and GPU:
| POCKET-35B IQ1_M | Bonsai-27B Q1_0 | ||
|---|---|---|---|
| CPU generate (Xeon, 16t) | 27.0 tok/s | 10.1 | ๐ข 2.69ร |
| GPU generate (H100) | 197 tok/s | 89 | ๐ข 2.22ร |
| GPU prompt (H100) | 753 | 1816 | ๐ด 0.41ร |
| Quality (HellaSwag, 400q) | 61.0% | 60.0% | โช tie (CI overlaps) |
[measured on a MacBook M3 Pro, 18 GB] โ and on a laptop, POCKET wins every axis, including prompt processing:
| POCKET-35B IQ1_M | Bonsai-27B Q1_0 | ||
|---|---|---|---|
| Metal generate (tg64) | 25.4 tok/s | 12.8 | ๐ข 1.99ร |
| CPU generate (8 threads) | 13.8 tok/s | 4.4 | ๐ข 3.13ร |
| Metal prompt (pp128) | 240.7 tok/s | 73.4 | ๐ข 3.28ร |
| CPU prompt (pp128) | 45.5 tok/s | 9.6 | ๐ข 4.75ร |
On a laptop GPU the arithmetic headroom that let Bonsai win prefill on an H100 is gone, so MoE sparsity wins across the board. POCKET-35B-Q2_K runs on the M3 Pro's CPU at 19.5 tok/s โ on an 18 GB Mac, run Q2_K on CPU (-ngl 0); its 13 GB exceeds the recommended Metal budget.
[measured โ GPQA Diamond, 198q, greedy] reasoning quality vs quantization:
| Model | GPQA-Diamond (greedy) |
|---|---|
| Qwen3.6-35B-A3B | 73.2% |
| POCKET-35B Q4_K_M | 68.7% |
| POCKET-35B Q2_K | 60.1% |
[pending โ community reports welcome] on-device iPhone and Strix Halo throughput. We publish only what we ran ourselves; help us fill the rest.
The same-size rival
Ternary-Bonsai-27B-Q2_0(7.2 GB) fails to load in upstream llama.cpp โ it needs the PrismML fork. POCKET runs on the tools you already have.
Files in this repo
| File | Size | Runs on | Korean PPL |
|---|---|---|---|
POCKET-KR-IQ2_M.gguf โญ |
5.1 GB | Android 8 GB+ | 7.95 |
POCKET-KR-160-Q2_K.gguf |
8.5 GB | phone/PC | 6.88 (quality-first) |
POCKET-KR-160-Q3_K_M.gguf |
11 GB | PC | 6.35 |
Built with our proprietary Korean-specific tuning. Active parameters unchanged โ same speed, half the size.
Quickstart
llama-cli -m POCKET-KR-IQ2_M.gguf -p "๋ํ๋ฏผ๊ตญ์ ์๋๋" -ngl 0 -t 8
On Android: PocketPal โ import GGUF.
Lineage โ where POCKET comes from
POCKET is quantized from Darwin-36B-Opus, VIDRAFT's flagship โ a model bred and evolved over several generations on the Darwin platform (crossbreeding, healing, expert surgery). Darwin-36B-Opus itself traces back to a Qwen3.5-family MoE architecture.
| Component | Origin |
|---|---|
| Starting checkpoint | Darwin-36B-Opus โ VIDRAFT, multi-generation Darwin evolution |
| Base architecture | Qwen3.5-family MoE (256 experts, top-8), unchanged |
Quantization (Q4_K_MโฆIQ1_M) |
stock llama.cpp โ no custom format |
| Runtime | upstream llama.cpp / Apple MLX โ unmodified |
| Proprietary language-specific tuning (KR/EN builds) | ours (VIDRAFT) |
The CPU/GPU speed comes from the sparse-MoE architecture plus ordinary quantization โ reproducible with the same base and the same tools. What we add is the Darwin-evolved weights, the honest measurement, the Korean tuning, and the pruning that makes the 5 GB phone builds.
Limitations
- The iPhone/Mac speed is not yet measured by us โ community reports welcome.
- Extreme quants (
IQ1_M) hurt Korean ~2.8ร more than English; useQ2_Kor larger for quality. - English phone builds trade quality for size; the PC build (
PC-mix) is much closer to full quality.
License
Apache-2.0.
POCKET is a VIDRAFT model family. 35B, in your pocket. No GPU.
Learn more
- On-device LLMs without a GPU โ and how POCKET measures up: Can you run a large LLM without a GPU?
- What model quantization is, and why a 4-bit model stays smart: What is model quantization?
๐งฉ The POCKET Family โ On-device AI by VIDRAFT
Big models, small hardware. No GPU, no cloud.
Models
- ๐ฆ POCKET-35B-GGUF โ flagship, PC / server, no GPU
- ๐ฆ POCKET-26B-GGUF โ compact 26B
- ๐ฐ๐ท POCKET-KR-GGUF โ Korean, Android
- ๐ POCKET-KR-MLX โ Korean, iPhone / Mac
- ๐ POCKET-EN-GGUF โ English, phone / PC
- ๐ป POCKET-Qwen3.8-Flash-Next-GGUF โ 180B on a laptop (8 GB VRAM + 32 GB RAM)
- ๐ผ๏ธ POCKET-Image-Zimage โ character-perfect text in any image
- ๐ฅ๏ธ POCKET-Zimage-CPU โ photoreal images on a CPU only
Demos & tools (Spaces)
- ๐จ POCKET-Image Studio โ text-in-image, generate in-page
- ๐ฅ๏ธ POCKET-35B-CPU โ 35B answering on a CPU
- ๐ฅ๏ธ POCKET-26B-CPU โ 26B on a CPU
- ๐ผ๏ธ POCKET-Zimage-CPU โ image generation on a CPU