Dual 3090 PC weighs local agents

The practical answer is less about the biggest model and more about keeping latency, context, and tool-calling reliable enough for daily coding.

• –Dual 3090s give 48GB VRAM, which is strongest for 30B-35B class coding models at higher quantization or 70B-class models at more painful speed/quality tradeoffs
• –vLLM is the better default when throughput, batching, long context, and OpenAI-compatible serving matter; llama.cpp still wins for GGUF simplicity, hybrid CPU/GPU offload, and quick experimentation
• –Qwen3.5/Qwen3.6 27B-35B class models are the likely sweet spot for local coding agents, while 100B+ “orchestrator” setups risk becoming slow demos unless the workflow tolerates latency
• –Agent harness choice matters as much as model choice: OpenCode, Cline-style flows, Claude Code-compatible adapters, and local OpenAI APIs are where the stack either becomes productive or collapses into prompt fiddling
• –Adding Whisper or whisper.cpp for local STT makes sense, but it is secondary to nailing inference stability, context length, and tool-call correctness first