Chat demos mislead: an agent holds a bloated context — tool outputs, file contents, its own plans — so the VRAM math that matters is weights + KV cache at your real context, not the benchmark card. Pick your setup below and see the numbers.
no signupruns entirely in your browsertuned for agentic workloads
—
Weights
KV cache
Runtime overhead
Estimates: weights ≈ params × bits-per-weight adjusted for embedding-heavy layouts; KV cache ≈ 2 × layers × kv-heads × head-dim × context × bytes (per the GQA shapes of current open models); overhead ≈ 1.2–2 GB for CUDA/llama.cpp buffers. Treat ±10% as noise, and verify with your own tps numbers.
The planner tells you what fits. Hermes Agent: 58 Prompts for Power Users tells you what to do with it — model picks per GPU tier, KV quant settings, and the agent workloads that justify each trade.