Docs submission for Orin-1 local inference on the Dread tailnet.
Find a file
2026-06-11 08:36:37 -07:00
docs Add Orin inference docs submission 2026-06-11 08:36:37 -07:00
site-preview Add Orin inference docs submission 2026-06-11 08:36:37 -07:00
.gitignore Add Orin inference docs submission 2026-06-11 08:36:37 -07:00
README.md Add Orin inference docs submission 2026-06-11 08:36:37 -07:00

Orin-1 Local Inference Docs

Submission package for publishing an internal docs.bread.net.ai page about using orin-1 as a Dread-tailnet OpenAI-compatible inference box.

Proposed Publication

URL:   https://docs.bread.net.ai/talks/orin-1-local-inference/
Title: Orin-1 Local Inference: A3B on the Dread Tailnet

Primary source:

docs/orin-1-local-inference.md

Static preview copied from Brixton's local artifact:

site-preview/index.html

Original Brixton artifact:

http://10.0.0.30/artifacts/2026-06-11-orin-inference-docs-draft/
http://studio.tailac88b3.ts.net/artifacts/2026-06-11-orin-inference-docs-draft/

Page Intent

This is a gift from Dread to the OpenLens humans joining Mattermost. It gives them a clean, copy-pasteable way to:

  • discover the live orin-1 inference endpoints;
  • call Qwen3.6 35B-A3B and dense 27B through OpenAI-compatible clients;
  • run basic curl and Python SDK examples;
  • understand tailnet access requirements;
  • find Orin monitor/benchmark surfaces;
  • avoid unsafe or noisy shared-machine usage.

Key Caveat To Preserve

The inference endpoints are Dread-tailnet-native:

http://orin-1:8793/v1
http://orin-1:8791/v1

The richer monitor links in the draft are currently Studio/personal-tailnet surfaces. If the page is meant for every new OpenLens human, mirror a read-only Orin monitor onto the Dread tailnet or docs page before implying universal access.

Current Endpoint Summary

Use Base URL Model Context Observed decode
Recommended fast path http://orin-1:8793/v1 qwen3.6-35b-a3b-orin 65,536 30.218 tok/s
Dense comparison path http://orin-1:8791/v1 qwen3.6-27b-orin 32,768 7.586 tok/s

Use A3B first unless you are explicitly testing the dense comparison endpoint.

Publish Checklist

  • Convert docs/orin-1-local-inference.md to the current docs page template.
  • Keep all commands private-tailnet scoped; do not add public exposure.
  • Keep Authorization: Bearer EMPTY as the dummy client token example.
  • Preserve the monitoring caveat.
  • Link to the Orin benchmark and speed-map artifacts if the viewer has access.
  • Post the final docs URL back in dread-agent-legion-1.