Skip to content

profiles.md

Latest commit

 

History

History
64 lines (50 loc) · 3.07 KB

profiles.md

File metadata and controls

64 lines (50 loc) · 3.07 KB
Raw

Profiles — the two ways to run hermes-max

You think in two profiles. They are the spine of the system. The six fine-grained modes are an advanced layer underneath; the profiles map onto the two you'll actually start with.

Profile A — Bring-Your-Own-GPU Profile B — No-GPU
For DGX Spark, Jetson Thor, RTX 6000/5090/4090, Mac Studio laptops, mini PCs, Mac minis, VPSes
Who drives (executes) your local model (free, sovereign, big context) DeepSeek V4-Flash over API
Who plans Kimi K2.6 (free, OpenRouter, 1M ctx) DeepSeek V4-Pro over API
Marginal cost ~$0 (electricity only) ~$17/month, no rate limits
Command hm up --free hm up --full
Maps to mode free full

Profile A — Bring-Your-Own-GPU (hm up --free)

You own a capable accelerator, so the heavy, high-volume work — every execute turn, the long-horizon grind, deep research distillation — runs locally and free. A free OpenRouter model (Kimi K2.6, 1M context) does the expensive planning, where a near-frontier model earns its keep.

  • Near-zero marginal cost. The only spend is optional free-tier accelerators (Groq/Cerebras), which cost $0.
  • Sovereign. Your code never leaves the box for execution.
  • Free uplift (optional). Deposit $10 on OpenRouter once to unlock 1000 free requests/day per model, then hm up --free --free-uplift adds a per-file coherence check at $0. See architecture.md §18.
  • Step up cheaply. Want stronger architectural judgment than Kimi-free? hm mode full-local swaps the planner to DeepSeek V4-Pro for ~$1.50/month while execution stays local and free.

Hardware floor and model choice: see hardware.md. The honest recommendation is a 24–32B-class local executor; below that, lean on cloud (Profile B) or heavier cloud uplift.

Profile B — No-GPU (hm up --full)

No capable GPU? The economic API path drives and plans entirely over the network. DeepSeek V4-Flash executes (hundredths of a cent per turn, 1M context, cache), V4-Pro plans. Both run through the funded DeepInfra endpoint by default, with direct DeepSeek as a configurable alternative.

  • ~$17/month for daily coding use, no rate limits — roughly 10% of a Claude Code Max subscription.
  • Runs on anything. A Mac mini, a $300 mini PC, or a small VPS is enough; the embedding/rerank services are optional and RAG falls back to BM25 automatically (see deployment.md).
  • The free accelerators (Groq/Cerebras/OpenRouter) still slot in for research fan-out and synthesis at $0.

Switching live

Profiles aren't a one-time choice. hm mode <name> reassigns every role chain without a restart, and hm mode --list shows all six postures. Start on the profile that matches your hardware; move between modes as the task warrants (e.g. frontier for one genuinely hard session, then back to full-local).

See modes.md for every mode and its role chains, and cost.md for what each dollar buys.