Strata runs a 125B model on a 12 GB gaming GPU via system RAM

The open-source Strata project runs Qwen3.8-Flash-Next, a 125-billion-parameter model, on gaming PCs with at least 12 GB of VRAM and 32 GB of RAM by loading 35–55 GB into system memory. It offers one-click installs for Windows and Linux and serves OpenAI- and Anthropic-compatible endpoints on localhost.

• Requires 12 GB+ VRAM and 32 GB+ RAM, loading 35–55 GB into memory
• Hits 94 tokens/s at Q2_0 and 53 tokens/s at IQ3_S on an RTX 5070
• The Coder variant reaches 91% of the full model's SWE-bench Verified score
• First message reads at about one minute per 30,000 tokens of context