Meridian Air CivoE

Meridian Air CivoE is webAI's Civilization of Experts system published on 2026-11-05 at https://www.webai.com/research/meridian-1-civoe (also announced on X by David Stout et al.).

Architecture:

• Meridian Air: a ~12B compact foundation, shipping in the webAI app from 2026-11-04;
• CivoE: task experts (documents, images, PII, contracts, prompt-injection resistance, …) share that fixed foundation — capability grows by specialization, not by enlarging the base;
• Self-reported on a curated suite from 14 public benchmarks: foundation alone 59.8 → CivoE 85.5; statistically tied with Gemini 4 (84.7) via Cursor agents, ahead of the tested Claude Opus 5.5 / GPT-5.6 setups.

The company and TwIL-8B are already tracked; this entry covers the next Meridian / CivoE product line. Meridian-1 LoRAs also appeared on HF around 11-02.

Claim

Instead of only scaling the foundation, webAI argues a fixed 12B plus experts can approach frontier generalists on selected specialized workloads. The page stresses this is not "small models replace frontiers everywhere".

Security narrative

CivoE separates reading untrusted text, critique and final answering; self-reported attack success on BIPIA / AgentDojo / InjecAgent drops sharply vs the lone foundation.

Caveats (also stated by webAI)

• Curated internal suite, not a public mega-leaderboard;
• Frontier models accessed via Cursor agents without fully matched revisions/settings;
• On-device privacy/cost economics not yet proven in matched experiments; CivoE rolls out over following weeks.

Who it's for

• Teams exploring local/edge small models + expert routing;
• webAI / TwIL users tracking Meridian;
• Security engineers studying layered prompt-injection defenses.

Getting started

1. Read https://www.webai.com/research/meridian-1-civoe ;
2. Watch Meridian Air / CivoE rollout inside the webAI app;
3. On HF webAI-Official, follow Meridian-1 LoRAs and Air quantization experiments.