DeepSeek-V4.1-Flash runs on 64GB of RAM, but commenters ask about VRAM first

Niko Veit tweeted that DeepSeek-V4.1-Flash can now run on 64GB of RAM with Strata and @mhelcig’s GSQ-RCO quantization. The tweet includes a video of a blue halo, but gives no speed, context length, or GPU requirements.

After the news came out, the questions in the comments were highly focused: how much VRAM is needed. Some asked about benchmarks for a 5090 with 64GB of RAM, some asked how fast an RTX PRO 6000 with 128GB of RAM could run it, and others asked whether 8GB or 10GB of VRAM would be enough. Others asked whether DGX Spark, Intel GPUs, and the 7900 XTX are supported under Windows. For now, those replies are still just questions waiting for results; there is no consensus answer.

Some users have already started installing it. The screenshot shows Strata’s launch screen, Engine version 0.1.48 cuda, model download progress at 61%, with 159.0GB downloaded out of 256.8GB. That size makes clear that being able to run on 64GB of RAM is one thing; preparing the disk and waiting for the download is another. Some also joked that running a Flash model on 64GB of RAM is pretty wild, and that the laptop fan has already started writing its resignation letter.

Another user, SmokeDev, posted the specs of a test machine: AMD Ryzen 7 5800X, 128GB RAM, RTX 3090 24GB, and plans to try it on this setup. This helps separate the two numbers: 64GB is the system RAM requirement, not the VRAM requirement. In the comments, some plan to build a machine with a 4080S and 64GB or 128GB of DDR4, while others warn it may be slow.

There is skepticism, too. User bruce says the model and its quantized version should be compared with a 27B Qwen; in his testing, performance was worse and it was unusable for everyday coding tasks. That claim needs more benchmarks to verify, but it at least shows there is still a gap between being able to start and being able to get work done.

So far, the tweet tells us this: there is now a way to run DeepSeek-V4.1-Flash on 64GB of RAM, and installation screenshots and hardware tests are starting to appear. What’s missing is data on speed, quality, VRAM usage, and compatibility across platforms. In the comments, some people are excited, some have already started installing it, and others are asking for benchmarks first. For local deployment, that order seems pretty normal.