V100 Benchmarks Favor Command R

Hot take: this is one of the more useful local LLM rig writeups because it is honest about what V100s can and cannot do instead of pretending legacy hardware behaves like Hopper.

• –Dense FP16 is the right default here; bitsandbytes 4-bit is the fallback when model size, not speed, is the constraint.
• –The benchmark spread says model architecture matters as much as raw parameter count: Command R 32B is materially more efficient than Gemma 4 31B or Qwen 2.5 72B on this stack.
• –The post is strongest when it becomes a compatibility map for Volta: vLLM runs, but modern optimization paths like FP8, FlashAttention2, and DeepSeek MLA are not the right target.
• –For legal workflows, the server is well-positioned for private summarization, extraction, drafting, and pattern recognition, but not for chasing the newest frontier-model features.
• –The writeup would be even better with a standardized prompt suite, batch-size disclosure, and separate warm/cold cache runs so the throughput numbers are easier to trust.