RTX 3090 still clears local LLM bar

For local LLM hobbyists, the 3090 is old but still one of the most defensible entry buys because VRAM matters more than raw gaming prestige. The catch is that the card opens the door to strong open models, not magic “local Claude” parity.

• –The big win is the 24GB frame buffer: that is enough for solid Qwen-class coding models, many 14B models at higher quality, and some 30B+ models once you lean on quantization.
• –The ceiling is real: 70B-class models, larger reasoning models, and long-context setups quickly force heavy quantization, CPU offload, or painful slowdowns.
• –The rest of the system matters more than newcomers expect; 24GB of system RAM is workable but not generous for larger model files, bigger contexts, and tooling overhead.
• –Hardware fit is not trivial: NVIDIA’s official 3090 specs call for a huge 3-slot card with 24GB GDDR6X, so PSU headroom, case clearance, thermals, and used-card condition matter almost as much as price.
• –At $699, the question is less “can it run local LLMs?” and more “is this particular used 3090 healthy enough to justify the power draw, heat, and noise versus waiting for a newer 24GB-class card.”