Use free models in the Zed editor with InferAll

Zed's agent can talk to any OpenAI-compatible endpoint. InferAll is one, so you can point Zed at it and use our free models ($0 input, $0 output) with one key.

1. Add the provider to settings.json

Open your Zed settings (zed: open settings) and add:

{
  "language_models": {
    "openai_compatible": {
      "inferall": {
        "api_url": "https://api.inferall.ai/v1",
        "available_models": [
          {
            "name": "nvidia/nemotron-3-ultra-550b-a55b",
            "display_name": "Nemotron 3 Ultra (InferAll, free)",
            "max_tokens": 131072,
            "capabilities": { "tools": true }
          },
          {
            "name": "openai/gpt-oss-20b",
            "display_name": "gpt-oss-20b (InferAll, free)",
            "max_tokens": 131072,
            "capabilities": { "tools": true }
          }
        ]
      }
    }
  }
}

max_tokens is the context budget Zed uses for the model. 131,072 is a safe value for both.

2. Give Zed your key

Get a key at inferall.ai/keys. Then either:

• enter it in Agent Settings (agent: open settings), which stores it in your system keychain, or
• set the environment variable Zed derives from the provider id above: INFERALL_API_KEY.

Do not put the key in settings.json.

3. Pick the model in the agent panel

Choose Nemotron 3 Ultra (InferAll, free) in the model menu and start a thread.

What we tested (2026-10-21)

We did not run the Zed app itself. We sent the request Zed's OpenAI-compatible provider sends (streaming, max_completion_tokens, a tool definition, tool_choice: auto) to InferAll, twice per model:

Model	Result	Time
nvidia/nemotron-3-ultra-550b-a55b	streamed, called the tool	1.1s, 1.5s
openai/gpt-oss-20b	streamed, called the tool	9.8s, 10.1s
z-ai/glm-5.3	streamed, called the tool	9.6s, 15.9s

A Zed user has already run the agent through InferAll this way (25 successful calls on Zed 1.24). Nemotron 3 Ultra was the fastest in our test, so we suggest it first.

If a free model is busy, InferAll answers from another free model and says which one in the response. Paid models (Claude, GPT, Gemini) work the same way with a funded key: add them to available_models with their prefixed ids, for example anthropic/claude-sonnet-4-6.

InferAll is an LLM aggregator: one API key for free open models and paid OpenAI, Anthropic, Google and xAI models. New accounts get 25 free calls on open models, no card needed.