# You Can Run DeepSeek V4 Flash on a Single RTX 3090. The Catch Is RAM, Not the GPU > Source: > Published: 2026-09-01 13:00:00+00:00 When [DeepSeek V4 Flash](https://vettedconsumer.com/deepseek-v4-flash-tested-frontier-class-coding-for-79-cents-a-day-and-it-runs-on-a-128gb-box/) landed, the takeaway was that a 284-billion-parameter model finally fit on a single high-memory box. DigitalSpaceport pushed that further: he ran the full 284B model on a single used RTX 3090, a $900 card, plus system RAM. It works, it is genuinely usable for chat, and the catch is not the GPU at all. It is the RAM. ## How a 284B model fits on a 24GB card The tool is a runtime called FreeToken (desktop apps for Windows, Ubuntu, and Arch, plus an image you can run on CasaOS). It does the same thing our [MoE-era piece](https://vettedconsumer.com/every-frontier-open-model-is-a-moe-now-what-that-does-to-your-hardware-math/) described: it keeps the model's rarely-touched expert weights in cheap system RAM and puts the hot, always-active parts on the GPU. Because DeepSeek V4 Flash is a Mixture-of-Experts with only about 13B active parameters per token, the GPU only ever handles a small slice, so a 24GB card can carry a model many times its size. As the reviewer put it, "it's going to be best when you're using MoEs." ## The measured speed, and the real bottleneck On the single 3090, DeepSeek V4 Flash generated **about 10 to 11 tokens per second**, varying with which experts fired for a given prompt. His read: "interactive speed... incredibly decent for chat. I don't think you're going to be running agentic stuff off of this." That matches the pattern from our [prompt-processing explainer](https://vettedconsumer.com/prompt-processing-vs-generation-why-your-box-is-fast-at-one-and-slow-at-the-other/): fine for a conversation, too slow for long agent loops. The number that actually gates this build is memory. He was blunt that a lot of "run it on one GPU" guides skip the dependency: "yes, one GPU is going to perform insanely well, but you have to have the RAM." His findings: | Factor | What he found | |---|---| | GPU | A single RTX 3090 (24GB) is enough | | System RAM for V4 Flash | 128GB "not enough"; 156GB works, 192GB is comfortable | | RAM speed | Matters a lot: DDR4-2400 is a real bottleneck vs faster kits | Because the expert weights stream from system RAM on every token, RAM bandwidth sets your ceiling once the GPU part is handled. He noted DDR4-2400 leaves speed on the table versus 3200, and a viewer agreed from the other side: "I knew there was a good reason I paid all that money for DDR5 6000." In the current market, that RAM is the expensive part of the build, not the used GPU. ## What viewers are saying - "I'm trying to convince my son he needs a 9070xt so dad can crunch tokens with his 'old' 3090." *@Cyber_Kel on YouTube* - "This is amazing! We are SO CLOSE to being able to run decent coding models on local hardware at good speeds. Wonderful times ahead!" *@devmindscapetutorials on YouTube* - "Qwen3-Coder 80B A3B q4km or q8km will be very useful on FreeToken." *@arshan2008 on YouTube* ## The takeaway This is the cheapest credible path to a genuinely frontier-adjacent model at home: a used 3090 you may already own, plus a big pile of system RAM. The scorecard is that it delivers chat-grade speed (10 to 11 tok/s), not agentic-coding speed, and the RAM you need (156GB or more, faster is better) now costs more than the GPU. If your workload is conversation, research, and the occasional code snippet, it is a remarkable amount of model for the money. If you need snappy agent loops, this is the wrong tool, and the buy-vs-rent math in our [cost calculator](https://vettedconsumer.com/cost-calculator/) will tell you when a rented node wins. Either way, check the exact RAM and quant your machine needs in our [Can I run it? calculator](https://vettedconsumer.com/can-i-run-it/) first. ## Sources and how we researched this - Video: ["FreeToken DeepSeek V4 Flash on a Single 3090 Local AI Testing"](https://www.youtube.com/watch?v=CrUabHJvPUM&ref=vettedconsumer.com)by DigitalSpaceport, August 24, 2026. All measured figures are his; we have not tested this build first-hand. - Model and architecture context: our [DeepSeek V4 Flash coverage](https://vettedconsumer.com/deepseek-v4-flash-tested-frontier-class-coding-for-79-cents-a-day-and-it-runs-on-a-128gb-box/)and the[MoE expert-offload explainer](https://vettedconsumer.com/every-frontier-open-model-is-a-moe-now-what-that-does-to-your-hardware-math/). - Viewer comments quoted verbatim and attributed from the video. *Related: **DeepSeek V4 Flash, tested** · **Every frontier open model is a MoE now** · **The used RTX 3090 for local AI** · **Prompt processing vs generation*