Bring your own model.
Choose how it behaves.
Roundtable AI loads any .gguf file — so the personality, the speed, and the guardrails are yours to pick. This is a plain guide to choosing one, from balanced mainstream models to faster small ones to community fine-tunes with fewer restrictions.
When it runs on your Mac, you choose the rules
Cloud assistants ship one personality and one set of guardrails, decided for everyone. A local model is a file you control — which means the debate can be as measured or as unfiltered as you want it.
Most hosted AI products are tuned for the widest possible audience: cautious, agreeable, and quick to decline. That is the right call for a mass-market chatbot — and exactly why a roundtable of them can feel like a panel that all secretly agrees.
Running a model locally changes the equation. You pick which model sits in each seat. Want a careful, well-behaved panel? Load a mainstream instruct model. Want debaters that actually take hard positions and push back without hedging? Load a community fine-tune with fewer guardrails. The choice — and the responsibility — sits with you, on your hardware, with nothing leaving your Mac.
A spectrum of GGUF models
Every one of these loads the same way in Roundtable AI: download the .gguf, point the app at it, assign it to an agent. The only difference is how the model behaves.
| Tier | Good for | Representative models | Pick |
|---|---|---|---|
| Balanced / mainstream | A reliable, well-mannered panel. The safe default for most discussions. | Llama 3.x Instruct, Qwen 2.5 Instruct, Mistral 7B Instruct, Gemma 2 | Default |
| Smaller / faster | Tight on RAM, or want more agents at once. Quicker, lighter, still capable. | Qwen 2.5 0.5B–3B, Llama 3.2 1B/3B, Phi-3 Mini | Lightweight |
| Fewer guardrails (community fine-tunes) |
Debaters that take strong positions and rarely refuse. Spicier, more opinionated exchanges. | Dolphin (uncensored fine-tunes), "abliterated" variants of Llama / Qwen / Gemma, and similar community releases | Advanced |
Finding GGUF models on Hugging Face
Hugging Face is the main hub for community GGUF models. These maintainers and searches are good starting points — links open Hugging Face directly.
Mainstream GGUF
Official and well-quantised builds of Llama, Qwen, Mistral, and Gemma. Search trending GGUF models or look for the Q4_K_M quant of a model you recognise.
Uncensored fine-tunes
The long-running Dolphin line of uncensored fine-tunes lives on cognitivecomputations. These are tuned to be helpful and direct rather than refusal-prone.
Fewer guardrails"Abliterated" variants
Abliteration is a technique that removes a model's refusal behaviour while keeping its capabilities. See mlabonne for the reference work, or browse all abliterated models and uncensored GGUF.
AdvancedTip: you do not have to commit to one tier. A great Roundtable AI session often mixes them — a balanced moderator keeping order while two fewer-guardrails debaters actually go at it.
About community models
Third-party models, your responsibility
The community models and links above are not created by, affiliated with, or endorsed by HighRoad Software. Uncensored and abliterated models have reduced or removed safety filtering, which means they can produce content that mainstream models would decline. You are responsible for how you use them, for complying with each model's licence, and for following the laws that apply where you live. Roundtable AI simply loads the .gguf file you choose — it does not host, modify, or distribute these models.
Pick a model and start a roundtable
Roundtable AI is free on Mac. Download any GGUF model, point the app at it, and seat it at the table in under a minute.