UAI Orb router v0
Linear classifier on @cf/baai/bge-m3 (1024-d) embeddings. Classes: SOLO, COUNCIL, RELAY.
This is not a user-traffic model. Training prompts are synthetic templates plus the public golden/orb eval cases. Held-out accuracy on that labeled set is in weights.json (heldOutAccuracy). Production keeps FEATURE_ORB_ROUTER off. Live chat uses the Orb menu (SOLO / RELAY / COUNCIL / RACE) and Cloudflare Workers AI incumbents only.
Live roster (2026-09-05, user-selected modes)
G0 reached 14 @cf models. REST G1 vs Mistral did not pass for GLM-5.3-flash Korean constraints; chat still uses flash as the SOLO tool core by product override. Voice/Meeting stay on Llama 4 Scout.
| Surface | Model |
|---|---|
| Chat SOLO core | @cf/zai-org/glm-5.3-flash (tools + vision, 112k usable) |
| Chat agent / document / code core | @cf/zai-org/glm-5.3 |
| Reasoning satellite | @cf/deepseek-ai/deepseek-v4-pro-0813 |
| Coding satellite | @cf/moonshotai/kimi-k2.7-code |
| Long context | @cf/deepseek-ai/deepseek-v4-flash-0731 (112k usable) |
| Instant chat | @cf/meta/llama-3.1-8b-instruct-fp8 |
| Voice / Meeting | @cf/meta/llama-4-scout-17b-16e-instruct |
RACE is a user toggle, not a router class. Satellites have no tools. The streamed core is never replaced (H-01). Loser RACE calls still bill.
Live chat on iuniai.com (2026-09-05) ran the three multi-model modes against this roster: RELAY (Z.ai core + DeepSeek polish), COUNCIL (Z.ai core + DeepSeek/Kimi/NVIDIA review), RACE (Z.ai core + DeepSeek/Kimi/NVIDIA; Qwen satellite failed once; arbiter still appended; losers billed).
Training code: train-orb-router.mjs in this repo (same file as scripts/train-orb-router.mjs on GitHub). Embeddings are @cf/baai/bge-m3. Labels are design-hypothesis only.
Paper: uai-orb-v0-paper.pdf in this repo.
- Downloads last month
- -