H100 tier

The local frontier: planners, critics, heavy executors, and every escalation from the swarm.

Layout

Two slots per card-or-pair: a reasoner/critic (Nemotron Super 49B via NIM, or DeepSeek-R1 Distill-70B via vLLM) and a heavy executor (Qwen3 32B FP8 or Llama 3.3 70B AWQ). On a single 80GB card, either time-share the slots, or run Qwen3 32B FP8 with the /think toggle covering both roles acceptably.

Role in the economics

This tier replaces credit-metered frontier calls for most judgment work — plans, reviews, twice-failed escalations. Keep a true frontier model only for the hardest architecture calls and final review of large merges. vLLM continuous batching means the whole swarm's escalations run concurrently; size max_num_seqs for it.

Reasoning models get token budgets (8–16k) and the LOW-CONFIDENCE stop rule — deliberation is the local analogue of credit burn.

Keeping judgment local is a big part of the Savings story — please consider donating to help support this project.