A 35B model that out-runs the top 1-bit 27B — on a CPU, no GPU.
Sparse Mixture-of-Experts · ~3B active per token · same stock llama.cpp, side by side
34.66B total~3B active/token0 GPUQ2_K · 13 GB
starting…
Ask anything. This runs on CPU — first reply may take a moment.
What is POCKET?
POCKET is a family of on-device builds quantized from Darwin-36B-Opus, VIDRAFT's flagship. A 34.66B sparse-MoE model that wakes only ~3B parameters per token — so it runs on a phone, and on a PC with no GPU, using the tools you already have. On the same machine and the same stock llama.cpp, it generates 2.7× faster on CPU and 2.2× faster on GPU than the leading 1-bit 27B, at matched quality.
This Space runs the Q2_K build on an upgraded CPU (no GPU). Speed here reflects a shared CPU box — a dedicated mini-PC is faster. Apache-2.0 · a VIDRAFT model family.
🌳 Bonsai-27B 1-bit dense · runs first
📦 POCKET-35B sparse MoE · runs second
Type a prompt — Bonsai answers first, then POCKET on the same CPU.
What am I looking at?
The exact same prompt is sent to Bonsai-27B (the most-downloaded on-device model, a 1-bit dense build) and to POCKET-35B (our sparse Mixture-of-Experts, Q2_K) — on the same CPU-only box, the same stock llama.cpp. Bonsai runs first, then POCKET. Watch the tok/s on each side.
Why POCKET wins on generation: a dense 27B reads all of its weights every token (~3.5 GB); POCKET's MoE wakes only ~3B parameters (~0.7 GB). The weaker the hardware, the wider the gap.
Honest notes: sequential race, each model gets the full CPU. tok/s is llama.cpp's measured generation speed. On a shared box a dedicated machine is faster. Prompt processing (not shown) is one place Bonsai leads on very long inputs. Quality is a statistical tie. Bonsai is a trademark of its owners; shown here for factual comparison only.