Post
487
š± POCKET ā a 35-billion-parameter model that runs on your iPhone, and on your PC with no GPU
We're releasing POCKET, VIDRAFT's flagship Darwin-36B-Opus compressed for on-device use. No fork, no CUDA, no cloud ā it runs on stock llama.cpp. It's a sparse Mixture-of-Experts model (256 experts, only 8 active per token), so the file can be large while the work per token stays small. That's what lets a 35B model run on a phone, and generate fast on a CPU with no graphics card.
Measured (POCKET-35B IQ1_M vs Bonsai-27B Q1_0):
ā¢ CPU generate (Xeon, 16 threads): 27.0 vs 10.1 tok/s ā 2.69Ć faster
ā¢ GPU generate (H100): 197 vs 89 tok/s ā 2.22Ć faster
ā¢ GPU prompt processing (H100): 753 vs 1816 ā 0.41Ć (Bonsai wins this one ā MoE prefill wakes every expert, so sparsity stops helping there. We say so.)
ā¢ Quality (HellaSwag, 400 q): 61.0% vs 60.0% ā a tie (confidence intervals overlap)
On a real consumer laptop ā MacBook M3 Pro (18 GB) ā POCKET wins every axis, prompt processing included:
ā¢ Metal generate: 25.4 vs 12.8 ā 1.99Ć
ā¢ CPU generate: 13.8 vs 4.4 ā 3.13Ć
ā¢ Metal prompt: 240.7 vs 73.4 ā 3.28Ć
One more quiet fact: the same-size, quality-oriented rival Ternary-Bonsai-27B (7.2 GB) fails to load in upstream llama.cpp at all ā it needs the PrismML fork. POCKET runs on the tools you already have: LM Studio, Ollama, PocketPal, MLX.
š Full story (tech, measurements, recipes): https://huggingface.co/blog/FINAL-Bench/pocket
Models:
š¦ POCKET-35B-GGUF (PC / server, no GPU): FINAL-Bench/POCKET-35B-GGUF
š°š· POCKET-KR-GGUF (Android): FINAL-Bench/POCKET-KR-GGUF
š POCKET-KR-MLX (iPhone / Mac): FINAL-Bench/POCKET-KR-MLX
š POCKET-EN-GGUF (English phone / PC): FINAL-Bench/POCKET-EN-GGUF
š„ļø Live demo (answering on a CPU, no GPU): FINAL-Bench/POCKET-35B-CPU
š Collection: FINAL-Bench/pocket-models-6a618ee5d23eafb7e185a5c6
We're releasing POCKET, VIDRAFT's flagship Darwin-36B-Opus compressed for on-device use. No fork, no CUDA, no cloud ā it runs on stock llama.cpp. It's a sparse Mixture-of-Experts model (256 experts, only 8 active per token), so the file can be large while the work per token stays small. That's what lets a 35B model run on a phone, and generate fast on a CPU with no graphics card.
Measured (POCKET-35B IQ1_M vs Bonsai-27B Q1_0):
ā¢ CPU generate (Xeon, 16 threads): 27.0 vs 10.1 tok/s ā 2.69Ć faster
ā¢ GPU generate (H100): 197 vs 89 tok/s ā 2.22Ć faster
ā¢ GPU prompt processing (H100): 753 vs 1816 ā 0.41Ć (Bonsai wins this one ā MoE prefill wakes every expert, so sparsity stops helping there. We say so.)
ā¢ Quality (HellaSwag, 400 q): 61.0% vs 60.0% ā a tie (confidence intervals overlap)
On a real consumer laptop ā MacBook M3 Pro (18 GB) ā POCKET wins every axis, prompt processing included:
ā¢ Metal generate: 25.4 vs 12.8 ā 1.99Ć
ā¢ CPU generate: 13.8 vs 4.4 ā 3.13Ć
ā¢ Metal prompt: 240.7 vs 73.4 ā 3.28Ć
One more quiet fact: the same-size, quality-oriented rival Ternary-Bonsai-27B (7.2 GB) fails to load in upstream llama.cpp at all ā it needs the PrismML fork. POCKET runs on the tools you already have: LM Studio, Ollama, PocketPal, MLX.
š Full story (tech, measurements, recipes): https://huggingface.co/blog/FINAL-Bench/pocket
Models:
š¦ POCKET-35B-GGUF (PC / server, no GPU): FINAL-Bench/POCKET-35B-GGUF
š°š· POCKET-KR-GGUF (Android): FINAL-Bench/POCKET-KR-GGUF
š POCKET-KR-MLX (iPhone / Mac): FINAL-Bench/POCKET-KR-MLX
š POCKET-EN-GGUF (English phone / PC): FINAL-Bench/POCKET-EN-GGUF
š„ļø Live demo (answering on a CPU, no GPU): FINAL-Bench/POCKET-35B-CPU
š Collection: FINAL-Bench/pocket-models-6a618ee5d23eafb7e185a5c6