Ficelle
Models
Free LLMs, by model and by job.
Popular models you can call for free today, with where and under which id. Or ask for the kind of model the job needs: Ficelle picks a free model that can do it, and moves to the next one when it fails.
Popular free models
- Qwen (Alibaba) Qwen3.8 27B Qwen3.8 27B is free on Groq, Cloudflare Workers AI and OVHcloud. The exact model ids, each free tier's limits, and how to fail over when one runs out.
- Google DeepMind Gemma 4 Gemma 4 31B and 26B A4B are free on OpenRouter, the Gemini API, SambaNova, Requesty and Cloudflare. Exact ids, free-tier limits and failover.
- Google Gemini 3.8 Flash Gemini 3.8 Flash has a free tier on the Gemini API: a 1M-token context, tools and structured output. What the free tier allows, its terms, and how to fail over.
- NVIDIA Nemotron 3 Super NVIDIA Nemotron 3 Super 120B A12B is free on Requesty and Cloudflare Workers AI. Exact ids, free-tier limits and failover.
- NVIDIA Nemotron 3 Ultra NVIDIA Nemotron 3 Ultra 550B A55B is free on OpenRouter and Requesty. The exact model ids, free-tier limits, and how to fail over when one runs out.
- OpenAI gpt-oss-120b OpenAI's open-weight gpt-oss-120b is free on Groq, SambaNova and Cloudflare Workers AI. Exact ids, free-tier limits and failover.
Model types: one id per job
- ficelle/auto-tools Tool calling Free models that can call your functions, with probe-failed models excluded and automatic failover behind one OpenAI-compatible id.
- ficelle/auto-json JSON & structured output Free models that return valid structured JSON for extraction pipelines, checked by probe, with failover when one breaks.
- ficelle/auto-fast Fast, low-latency Free models for quick, low-stakes calls: titles, labels, one-liners. Latency-first routing with a short timeout so failover starts fast.
- ficelle/auto-reasoning Reasoning Free models that expose reasoning controls, checked by probe, for multi-step logic, planning and small code changes.
- ficelle/auto-coding Coding assistants Free models for coding assistants, restricted to a certified coding pool. Fails loudly instead of silently swapping in an uncertified model.
- ficelle/auto-orchestrator Agent orchestration Free models for the main loop of an agent: heavier multi-step work with tools and structured output, checked by an orchestration probe.
- ficelle/auto-long Long context Free models for large prompts: long documents, big threads, whole files, checked by a long-context probe.
- ficelle/auto-compression Context compression Free models that compact long conversation history so agents keep working inside smaller context windows.
- ficelle/auto-vision Vision Free image-understanding models behind one OpenAI-compatible id, checked with a real image probe, with failover across providers.
- ficelle/auto-audio Audio input Free models that accept audio input, checked with an audio probe, with failover across providers.
- ficelle/auto-video Video input Free models that accept video input, checked with a video probe, behind one OpenAI-compatible id.
- ficelle/auto-multimodal Multimodal Free models that take image, audio or video input, routed to a model whose modality checks have not failed.
What every type has in common
- Free pricing: a model with any price is never eligible, and paid fallback is hardcoded off.
- Tool calling support.
- At least 64,000 tokens of context by default.
Each type is a routing policy, not a model: it keeps a ranked pool of eligible free models across the providers you configure and fails over inside that pool. Pro adds custom virtual models built on top of these.
