Ficelle
Providers
Free LLM API providers, compared from the source.
What each provider really gives away for free, the limits they publish, and how to plug them into one OpenAI-compatible endpoint with failover. Every claim links to the provider's own docs, checked on 28 September 2026.
| Provider | What is free | OpenAI-compatible | In Ficelle | Checked |
|---|---|---|---|---|
| OpenRouter | :free model variants, 20 RPM, 50 or 1,000 requests/day | Yes | Core | |
| Nous | Subscription portal; some catalog models priced at 0 | Yes | Core | |
| Mistral | Free mode, no card; monthly credits under fair use | Partly | Pro | |
| Groq | Free plan, e.g. 30 RPM, 1K RPD, 8K TPM per model | Partly | Pro | |
| Gemini | Free tier on many Flash and Gemma models | Yes | Pro | |
| NVIDIA | Free prototyping on trial credits, not production | Yes | Pro | |
| Cerebras | No free tier; $5 trial credits for 30 days, card required | Partly | Pro | |
| SiliconFlow | Free models on siliconflow.cn; $1 credit on .com | Yes | Pro | |
| Requesty | Free plan, 200 requests/day and 20/min, no card | Yes | Pro | |
| OVH Cloud | Anonymous calls at 2 RPM per IP and per model | Yes | Pro | |
| SambaNova | Free tier, 20 RPM and 20 RPD per model | Yes | Pro | |
| Cloudflare | 10,000 Neurons per day, resets 00:00 UTC | Partly | Pro | |
| Routeway | :free models, 200 requests/day, low-priority queue | Yes | Pro |
Why one free tier is never enough
Every free tier above is metered: per minute, per day, per model or per account, and the numbers move without notice. A single key is a countdown to the next 429. Ficelle pools the providers you configure behind one local endpoint, routes only to free models, and moves to the next eligible model when one fails. Read why provider breakage kills agents, or pick a model type.
For help comparing quotas, tool support, privacy, and failover, read the free LLM API guide.
