VRAMCalculator.com

Can Your PC Run DeepSeek V4.1 Flash with Project Maya?

Setup

Graphics card

Unified device

Cards

System RAM

Context

System

Requests at once

Drafting

INT8 latent cache (a smaller KV cache, no measurable quality change) KV streaming (Maya's default) Pictures: the vision tower, 320 MB on the first card (Maya's default) Keep every expert in RAM (fails if it does not fit)

Layers split by each card's room. Each card holds its own layers, state and reserve; the last card carries the draft layer. Pools add up; RAM is shared.

Build	RAM	Fixed	Experts	Verdict

Closest measured setups

How this is calculated

GLM 5.3 Flash: 321B, 18B active. Hottest experts in VRAM, next in RAM, rest on the SSD. KV 12.7 KB/token FP16, 7.4 INT8. DFlash2 adds 0.7 GB. Speeds are the reporters' own.

More on this model: Project Maya with GLM 5.3 Flash.

The Local LLM Sizing Guide, Issue 1: what runs on October's new hardware

Issue 1 sizes GLM 5.3 Flash for the machines that hold it whole, from one DGX Spark up. And every other model, sized for your machine.

Know what fits before you buy

From $9 a month. Card or crypto. The October issue included.

Can Your PC Run DeepSeek V4.1 Flash with Project Maya?

Closest measured setups

What Maya reports for DeepSeek