Council seats
Models are swappable “seats”, not “the smartest models.” Defaults live in
council.config.yaml, are tracked in a dated MODELS.md + a machine-readable
registry, and concrete benchmark numbers are pulled live from leaderboards,
never frozen into the spec.
Online frontier council seats (via BYO keys / OpenRouter)
Section titled “Online frontier council seats (via BYO keys / OpenRouter)”State June 2026:
| Seat | Model | Provider |
|---|---|---|
| Reasoning | Claude Opus 4.8 | Anthropic |
| Allround | GPT-5.5 (Pro) | OpenAI |
| Multimodal / long-context | Gemini 3.1 Pro | |
| Cost / frontier | DeepSeek V4 or Qwen3.7 Max | DeepSeek / Alibaba |
| Chairman (default) | Gemini 3.1 Pro or Claude Opus 4.8 | Google / Anthropic |
Correction (Devil’s Advocate): Grok 4.1 was dropped. Grok 4.1 Fast is deprecated and shuts down August 2026. For a fourth “different perspective” seat, use another house (e.g. a current Mistral or Llama frontier model).
Local models (GGUF): verified options
Section titled “Local models (GGUF): verified options”| Seat / role | Model (GGUF, quantized) | License | ~Size (Q4_K_M) |
|---|---|---|---|
| Reasoning | DeepSeek-R1-Distill-Qwen-32B | MIT ✅ | ~20 GB |
| Agent / tools | Qwen3-32B | Apache-2.0 ✅ | ~20 GB |
| General (mid) | Qwen3-8B | Apache-2.0 ✅ | ~5 GB |
| Lightweight | Gemma 4 (E4B/4B class) | Gemma Terms ⚠ | ~3-5 GB |
| Vision (VLM) | Qwen-VL / Moondream (GGUF) | per model | ~3-8 GB |
| Voice (STT) | whisper.cpp large-v3-turbo / base | MIT ✅ | ~0.1-1.5 GB |
Which seats/roles are active is set in council.config.yaml;
redistribution rules differ per model. Read licenses before
bundling anything.