What is the best combo of free models for the council?

A council has four seats: three Thinkers and one Judge. The best free setup is rarely four copies of one model. It is a mix of model families from different providers, so the Thinkers disagree usefully and no single rate limit slows everyone down. This page suggests combinations for the most common jobs, then shows how to enter them in Settings.

Free model lineups change often. Names below were checked in October 2026, but a model can be retired or moved behind a paywall at any time. If a model is missing from your provider's list, pick the closest one in the same family.

The short version

Always keep at least two different providers in any council. Putting all four seats on one provider is the fastest way to hit a rate limit.

How to set it up in Settings

Open Model settings (the first time, you will see Connect your models). Each of the four seats has its own card: Thinker 1, Thinker 2, Thinker 3 and the Judge.

1. Choose a provider. Use the provider dropdown on each card. Next to it, a Get a free key ↗ link takes you straight to that provider's key page.
2. Paste the API key. Use the eye icon to show or hide it. Each provider has a usual key prefix, and if your key does not start with it, the card turns amber and asks you to double-check. That is almost always a key pasted from the wrong provider.
3. Pick the model. The model field is a text box with suggestions: tap it to see the list, or type any model ID yourself. Press Load live models under the card to fetch the provider's current list. This is the quickest way to confirm that a free model is still available. OpenRouter's list loads without a key, so you can browse it before signing up.
4. Set a specialty for each Thinker. Every Thinker has a Specialty dropdown (reasoning, coding, research and more), with a one-line description under it. The specialty tells that Thinker which expertise to answer with, so match it to the job in the tables below and give your three Thinkers different specialties when the task is broad. The Judge has no specialty, since its job is to merge the drafts.
5. Save. Save only becomes available once all four seats have a key and a model. Keys are encrypted and saved to your account, not to your browser, and are sent only to the provider you chose.

Things worth knowing

• One key can serve several seats. If two seats use the same provider, paste the same key in both. You need a key for each provider you use, not for each seat.
• Changing a provider resets that seat. Switching the provider dropdown clears the key and swaps the model to a suggested one for the new provider. Your specialty is kept. Choose the provider first, then the key and model.
• "Use Thinker 1's provider & key for all" is a shortcut, not a recommendation. It copies Thinker 1's provider and key to every seat and fills in suggested models, which is handy if you only have one key and want to start fast. It also removes the provider mix that makes a council work, so switch at least one or two seats to other providers afterwards.
• Vision warning. Providers whose models cannot read images show a note on the card saying that seat answers from text only. If you plan to send screenshots or charts, avoid those providers (see the images combo below).
• Custom provider. If you run your own endpoint or use a service that is not listed, choose Custom and enter its base URL (for example https://api.example.com/v1). Save stays disabled until the base URL is filled in.
• Clear saved keys. The red button at the bottom removes every saved key from your account after a confirmation. Use it before sharing your screen or if a key leaks, and then revoke the key at the provider as well.
• Required vs optional. On first setup the window cannot be dismissed until the council is complete. Afterwards you can press Escape, click outside it, or use Cancel.

Free models by provider

OpenRouter

The widest free menu. Model IDs ending in :free cost nothing. The strongest ones right now:

• Reasoning: NVIDIA Nemotron 3 Ultra and Nemotron 3 Super, both with very long context.
• Coding: Qwen3 Coder, Poolside Laguna, Cohere North Mini Code.
• General: Llama 3.3 70B, GPT-OSS 120B.
• Images: Gemma 4 31B and Nemotron 3.5 Super VL.

Limits are about 20 requests per minute, and 50 requests per day until you make a one-time $10 credit purchase, which raises the daily cap to 1,000. Put at most two seats here, and use the daily cap as a reason to spread seats across other providers.

Google Gemini

The most generous free tier and the best free option for reading images. Recent Flash and Flash-Lite models are free. Pro models have moved in and out of the free tier this year, so press Load live models to check what your key can actually call before assigning one to the Judge. Free-tier prompts can be used by Google to improve its products, so avoid pasting private material.

NVIDIA NIM

A large catalog of free hosted open models with a limit of around 40 requests per minute. Good picks include DeepSeek V3.2 and GLM-4.7 for reasoning and coding, and MiniMax M2.7 for coding and long tasks. This is the easiest place to find a strong Thinker that is not a Llama or a GPT-OSS.

Groq

Very fast, with limits that differ per model. Good free picks include GPT-OSS 120B and 20B, Llama 4 Scout, and Qwen3. Because the limits are per minute and per day, Groq works best as one or two seats.

Cerebras

Extremely fast, with around 1 million free tokens per day. The free tier has a short context window (about 8K, longer on request), and the model list has shrunk this year. GPT-OSS 120B and Llama 3.3 70B are the usual choices. Its models read text only, so Settings will show the text-only note on any Cerebras seat. Do not use Cerebras as the Judge, because the Judge has to read every Thinker's answer at once.

GitHub Models

Needs only a GitHub account. Limits are tight: roughly 15 requests per minute and 150 per day for lower-tier models, and far less for reasoning models such as DeepSeek-R1. It also offers OpenAI GPT models on the free tier. Use it for one seat, and not for a Thinker that needs a long reply.

Mistral

The free Experiment tier has a generous monthly token cap (around a billion tokens) but a low request rate. That suits a council well, since each seat only answers once per question. Use Codestral or Devstral for code and the current Large or Medium model for writing.

Hugging Face

A small monthly credit, so treat it as a spare seat rather than a main one. It is useful for trying a specific open model that the other providers do not host.

Custom (your own endpoint)

For any service not in the list. You supply the base URL, key and model ID yourself, so the limits and privacy rules are whatever that service sets. A good home for a self-hosted model or a provider that is new to the list.

Best combos in detail

For each combo, set the Thinkers' specialties to match the job and leave the Judge as is.

For reasoning (maths, logic, planning, hard questions)

• Thinker 1: NVIDIA NIM, DeepSeek V3.2
• Thinker 2: OpenRouter, Nemotron 3 Ultra :free
• Thinker 3: Groq, GPT-OSS 120B
• Judge: Gemini, the strongest free model your key can reach
• Specialties: reasoning on all three, or reasoning plus research on one if the question needs facts

Why it works: three different families (DeepSeek, NVIDIA, OpenAI-style open weights) tend to fail in different ways. Gemini as Judge adds a fourth view and the longest reliable context for reading all three answers.

For coding

• Thinker 1: OpenRouter, Qwen3 Coder :free
• Thinker 2: NVIDIA NIM, GLM-4.7
• Thinker 3: Mistral, Codestral or Devstral
• Judge: OpenRouter, Poolside Laguna :free, or Gemini if Laguna is rate limited
• Specialties: coding on all three, or reasoning on one to review the logic

Why it works: Qwen, GLM and Mistral code models are trained on different data, so they catch each other's bugs. Keep prompts focused on one file or one function, since free tiers cap output length.

For writing and everyday questions

• Thinker 1: Gemini, free Flash
• Thinker 2: Mistral, current free model
• Thinker 3: OpenRouter, Llama 3.3 70B :free
• Judge: Gemini, free Pro if available, otherwise Flash
• Specialties: the general or writing-oriented option if the list has one, otherwise research

Why it works: Gemini, Mistral and Llama write in noticeably different styles, which gives the Judge real choices to merge.

For images, screenshots and charts

• Thinker 1: Gemini, free Flash
• Thinker 2: OpenRouter, Gemma 4 31B :free
• Thinker 3: OpenRouter, Nemotron 3.5 Super VL :free
• Judge: Gemini, free Flash

Why it works: every seat can read images. Do not use Cerebras here, since its models only read text.

For speed

• Thinker 1: Cerebras, GPT-OSS 120B
• Thinker 2: Groq, GPT-OSS 20B
• Thinker 3: Groq, Llama 4 Scout
• Judge: Gemini, free Flash

Why it works: answers arrive in a few seconds. The trade-off is less variety, since two Thinkers share a family, and Groq's per-minute limits can bite if you ask many questions in a row.

Rules of thumb

• Give the Judge the most capable and longest-context model you have. It reads three full answers before it writes one.
• Never put the Judge on Cerebras or GitHub Models. Their free input limits are too small.
• Pick Thinkers from different families. Two copies of the same model mostly repeat each other.
• Use each provider on at most two seats. That keeps you under every free limit.
• Give Thinkers different specialties for broad questions, the same one for focused jobs. A coding question benefits from three coders. A product decision benefits from reasoning, research and writing views.
• Check the model list before saving. Press Load live models on each seat so you are not saving a model that was retired last week.
• Keep one backup key. If a free model is retired or rate limited, swap that seat to a spare provider instead of rebuilding the council.
• Do not paste private data into free tiers. Some free tiers, including Gemini's, may use prompts to improve their products.

Troubleshooting

• Save is greyed out. One of the four seats is missing a key or model, or a Custom seat has no base URL.
• The card is amber and mentions a key prefix. The key probably belongs to a different provider. Re-copy it from the provider's key page using the Get a free key ↗ link.
• "No models returned" or an error when loading models. The key may be wrong, expired or out of quota. Type the model ID manually and test with a short question.
• I changed a provider and my key vanished. That is intentional, so a key is never sent to the wrong provider. Paste the new key after choosing the provider.
• Answers are slow or fail on one seat. That provider is probably rate limited. Move that seat to a different provider and keep the rest.