VIDRAFT.
VIDRAFT / Insights / On-device AI
On-device AI

Can you run a 35B AI model with no GPU?

How VIDRAFT's POCKET Box and POCKET-35B run a large language model on-device, CPU-only, and how fast it is.

Published 2026-07-31About 2min readby VIDRAFT
Quick answer

Yes. VIDRAFT's POCKET Box is a CPU-only appliance with no GPU that runs POCKET-35B on-device, a 34.66B-parameter Sparse MoE model that activates only ~3B parameters per token. On a 16-thread Xeon CPU it reached 27.0 tok/s at IQ1_M (developer-measured under stated conditions, not third-party verified).

Can you run a large language model without a GPU?

Yes. POCKET Box is a CPU-only appliance with no GPU that runs POCKET-35B on-device, and its development just succeeded.

POCKET-35B is a Sparse MoE model with 34.66B total parameters that activates only ~3B per token, built on Darwin-36B-Opus (a Qwen3.5-family MoE). In-domain expert pruning cuts 256 experts down to 128, halving the model size. It runs on unmodified upstream llama.cpp and MLX with no runtime fork required.

Why does POCKET-35B run faster on CPU?

CPU inference is memory-bandwidth bound, and as a Sparse MoE, POCKET-35B reads only 0.66 GB per token versus 3.5 GB for a same-size dense model, so it moves far less memory.

Because it reads fewer weights per token, the weaker the hardware, the more the MoE wins. On a 16-thread Xeon CPU, POCKET-35B hit 27.0 tok/s at IQ1_M, 2.69x faster than the same-size 1-bit Bonsai-27B (10.1 tok/s). MoE-aware mixed-precision quantization compresses Korean harder and prunes English harder.

How fast is POCKET Box on different hardware?

Developer-measured: 27.0 tok/s on a 16-thread Xeon CPU, 197 tok/s on H100, and 25.4 tok/s via Metal (13.8 CPU) on an Apple M3 Pro 18GB.

On H100 it ran 197 vs 89 tok/s, a 2.22x edge over the same-size baseline. On mobile, iPhone's POCKET-EN ships as a 5.3GB mixed-precision GGUF and Android's POCKET-KR as a 5.1GB MLX 2-bit build, with 88% and 94% routing consistency respectively. All figures are self-measured under stated conditions and are not third-party verified.

How do you install POCKET-35B and which quantization should you use?

It is distributed via Ollama, Docker, ModelScope, and Hugging Face GGUF; use Q4 or higher rather than extreme low-bit like IQ1 to preserve Korean quality.

Because POCKET-35B runs on unmodified llama.cpp and MLX, you can use the standard toolchain directly. Extreme low-bit quantization such as IQ1_M is fast but degrades Korean output, so Q4+ is recommended for real use. The benchmark numbers are developer self-measured and provided for reference.

Frequently asked questions

Does POCKET Box need a GPU?
No. POCKET Box is a CPU-only appliance with no GPU at all.
Can it run on iPhone or Android?
Yes. iPhone's POCKET-EN is a 5.3GB GGUF and Android's POCKET-KR is a 5.1GB MLX 2-bit build.
Which quantization is recommended?
Q4 or higher is recommended for Korean quality; extreme low-bit like IQ1 is faster but degrades Korean output.

Sources

Related

AI for Science
Which AI model is best at predicting a drug's human intestinal absorption (HIA)?
Model reasoning
What is the best Korean LLM on the GPQA Diamond science benchmark?
AI safety
Does AI know what it does not know? What is FINAL-Bench?
↖ Home — vidraft.net

This article is based on VIDRAFT public, measured data and external sources. Performance figures are measurements under the stated conditions and may vary by environment.