NVIDIA NIM / Nemotron
NIM containers expose an OpenAI-compatible API, so all Anchor tooling works unchanged.
The reasoning toggle
Llama-Nemotron models switch modes via a system-prompt line: detailed thinking on (deliberate reasoning — temp 0.6/top_p 0.95) or detailed thinking off (direct answers — greedy). The fleet client (anchor_client.py) applies this automatically for endpoints with quirks: {think_toggle: nemotron, temperature: 0, sampling_thinking: {top_p: 0.95}} — greedy when thinking is off, 0.6/0.95 when it's on.
One model, three roles
This toggle gives you the orchestrator pattern on a single deployment — thinking on = judgment, thinking off = keystrokes:
| Role | Toggle | Addition to mythos-core |
|---|---|---|
| Planner | on | "Only output is a plan per the template; no implementation." |
| Executor | off | one task spec |
| Critic | on | "Review, don't fix; use the review template." |
Nemotron Super/Ultra with thinking on is the most credible local stand-in for a frontier planner/critic; Nano works as a spec executor. Don't leave thinking on for bulk execution — deliberate tokens are the local equivalent of credit burn.
Local judgment and execution are a big part of the Savings story — please consider donating to help support this project.
Tracked plans
Same .plans/ contract as other platforms: draft under drafts/, human promotes to ready, claim → in-progress/, finish → completed/. Docs describe current state, not plans — never write product docs from .plans/ contents; document shipped code only.