A Curated Map of Free LLM Inference and Fine Print

What it does This repository is a living directory of services that offer free or trial-credit access to LLMs via standard APIs. It splits providers into two camps—truly free tiers and those offering limited trial credits—while listing specific models, rate limits, and usage restrictions for each. The README is auto-generated by a Python script (src/pull_available_models.py), suggesting it attempts to stay current as offerings change.

The interesting bit The maintainer draws a hard line against sketchy reverse-engineered chatbot wrappers, focusing only on legitimate providers. That hard line makes it a useful filter in a space crowded with wrapper scams. The granularity is also unusual: instead of vague promises, you get exact request-per-minute caps, token limits, and warnings about data-training opt-ins.

Key highlights

• Covers major free providers (OpenRouter, Google AI Studio, Groq, Cerebras, Cloudflare Workers AI, etc.) and a separate list of trial-credit providers.
• Documents specific gotchas: Google AI Studio trains on your data outside the UK/EU/CH/EEA; Mistral’s free tier requires opting into data training; GitHub Models imposes extremely restrictive token limits tied to Copilot tiers.
• Rate limits are spelled out in provider-specific tables—e.g., OpenRouter allows 20 requests/minute and 50/day on the free tier, while Cerebras allows 14,400 requests/day for some models.
• Explicitly excludes illegitimate services that reverse-engineer existing chatbots.

Caveats

• The list is only as current as the last script run; some provider offerings may have shifted since generation.
• Several free tiers require phone-number verification (NVIDIA NIM, Mistral) or come with tight quotas that make them impractical for anything beyond prototyping.
• Trial-credit providers are mixed in with truly free ones, so read the section headers carefully before you commit to an integration.

Verdict Worth bookmarking if you are prototyping, experimenting, or just refuse to pay for tokens before validating an idea. Skip it if you need production-grade reliability or predictable costs—most of these tiers are guard-railed hobbyist lanes.