// the take
A reference list of LLM providers offering free-tier text inference APIs, split into model-training companies (Google, Mistral, Cohere, etc.) and third-party inference hosts (Groq, OpenRouter, Cloudflare Workers AI, etc.). It's for developers prototyping agents or side projects who want to avoid paying for API access, with every entry giving context length, rate limits, and OpenAI SDK compatibility.
The footnotes do real diligence instead of repeating marketing copy — they call out exactly when a provider trains on your free-tier prompts, which free models are actually excluded from the stated free quota (Cloudflare's seven carved-out models), and concrete deprecation dates for specific models. There's an actual verification pipeline (scripts/verify-providers.js, .verify/ reports) that live-probes endpoints rather than trusting docs, and the Kilo Code entry is honest that its catalog lags what's actually being served. Maintenance looks active — retirement and shutdown dates are tracked per model, not just added once and forgotten.
"Permanent" free tier is doing a lot of work here: the footnotes themselves document provider after provider changing limits, retiring models, or adding new verification requirements without notice, so anyone building production reliance on this list is building on sand. Several of the more generous entries (ModelScope, SiliconFlow) require real-name/national-ID verification and can't even be tested outside mainland China, which is a real adoption blocker buried in a footnote rather than flagged where someone picks a provider. Most rate limits are too low for anything beyond prototyping (2 RPM anonymous, 1 concurrent request, 15 RPM/20K tokens per day) and the list doesn't help with the actual hard problem — failover when a free tier gets throttled or pulled — that's left entirely to the reader, with the repo's own sponsors (manifest.build, okapi.bot) implicitly selling the thing this list doesn't do.
We read the README, the directory tree and a key source file in November 2026. Nobody pays to be listed. How we judge →