LocalLLaMA asks best coding, vision model

The ask is less about one magic model and more about assembling the right stack for a 12GB card. On that hardware, the smartest setup is usually a compact reasoning model plus a separate vision-capable model, not one oversized all-rounder.

• –DeepSeek-R1-style reasoning models are strong for step-by-step thinking, but they can be heavy for a 12GB GPU unless heavily quantized or partially offloaded
• –Qwen-family coding models remain a common recommendation for local developer work because they balance code quality, tool use, and deployability
• –Vision needs are a separate constraint: multimodal models like Qwen2-VL or Llama 3.2 Vision are better fits than trying to force a text-only coder to handle images
• –With 64GB system RAM, the machine can spill into CPU memory, but latency still makes model selection matter more than raw capacity
• –The thread is a good snapshot of LocalLLaMA’s current consensus: optimize for a workflow, not a single benchmark winner