Qwen 27B marks a “DeepSeek” moment for open source, runs on a single RTX 5090

X user Chubby calls Qwen 27B a “DeepSeek” moment for open source, a model that reaches the performance of closed models from a few months ago and runs locally on a single RTX 5090. He says the model is comparable to GPT-5.6 Luna in Artificial Analysis tests, with the practical implication of cost savings versus cloud access.

The RTX 5090 currently sells for about $4,500 because demand exceeds supply. According to Chubby’s calculation, a Luna task costs $0.05, so buying the card is equivalent to roughly 90 thousand tasks, or about 60 tasks per day for four continuous years. Those figures assume full hardware availability and zero electricity or maintenance costs.

In the thread replies, a known limitation surfaced: the same VRAM required by Qwen 27B is also used by Codex and parallel development environments. Users running several tools simultaneously may find insufficient VRAM, making the “runs on 5090” claim dependent on the specific workload configuration.

The core argument emerging from the discussion is that local models need not win every benchmark; they only need to be “good enough” for privacy and full control to justify giving up cloud convenience. Chubby says Qwen 27B meets that threshold. However, the comparison to Luna rests on a single Artificial Analysis test rather than a broad independent evaluation, and real-world performance may vary with task type and settings.