The whole board
Every endpoint. Not a deduped list.
The same model on two providers is two different realities — different quotas, different signup friction, different data policy, different verification date. So this is a matrix of 117 provider × model endpoints, not 107 model names with the useful part averaged away.
36 providers carrying models · 81 free endpoints · every row traceable to a dated source on its provider page.
| provider | model | context | rpm | rpd | steps | free | trains? |
|---|---|---|---|---|---|---|---|
| GitHub Models | deepseek/DeepSeek-R1 | 64,000 | 15 | 150 | 2 | Yes | Unknown |
| GitHub Models | openai/gpt-4.1 | 1,000,000 | 10 | 50 | 2 | Yes | Unknown |
| Cerebras Inference | gemma-4-31b | — | 5 | — | 3 | Yes | Unknown |
| Cerebras Inference | gpt-oss-120b | 131,072 | 5 | — | 3 | Yes | Unknown |
| Cohere | command-a-03-2025 | 256,000 | 20 | — | 3 | Yes | Unknown |
| Cohere | command-a-plus-05-2026 | 128,000 | 20 | — | 3 | Yes | Unknown |
| Google AI Studio (Gemini API) | gemini-2.5-flash | 1,048,576 | 5 | 20 | 3 | Yes | Yes |
| Google AI Studio (Gemini API) | gemma-3-27b-it | — | 30 | 14,400 | 3 | Yes | Yes |
| Groq | llama-3.1-8b-instant | 131,072 | 30 | 14,400 | 3 | Yes | Unknown |
| Groq | llama-3.3-70b-versatile | 131,072 | 30 | 1,000 | 3 | Yes | Unknown |
| Hugging Face Inference Providers | deepseek-ai/DeepSeek-V3-0324 | — | — | — | 3 | Yes | Unknown |
| Hugging Face Inference Providers | meta-llama/Meta-Llama-3.1-8B-Instruct | 128,000 | — | — | 3 | Yes | Unknown |
| Ollama Cloud | deepseek-v3.1:671b-cloud | 131,072 | — | — | 3 | Yes | No |
| Ollama Cloud | gpt-oss:120b-cloud | 131,072 | — | — | 3 | Yes | No |
| OpenRouter (free models) | google/gemma-4-31b-it:free | 262,144 | 20 | 50 | 3 | Yes | Yes |
| OpenRouter (free models) | meta-llama/llama-3.3-70b-instruct:free | 131,072 | 20 | 50 | 3 | Yes | Yes |
| Z.ai (Zhipu AI) | glm-4.5-flash | — | — | — | 3 | Yes | Unknown |
| Z.ai (Zhipu AI) | glm-4.6v-flash | 128,000 | — | — | 3 | Yes | Unknown |
| Mistral La Plateforme | mistral-large-2411 | 262,144 | — | — | 4 | Yes | Yes |
| Mistral La Plateforme | mistral-medium-2604 | 262,144 | — | — | 4 | Yes | Yes |
| OpenCode Zen | big-pickle | — | — | — | 4 | Yes | Yes |
| OpenCode Zen | deepseek-v4-flash-free | — | — | — | 4 | Yes | Yes |
| Cloudflare Workers AI | @cf/meta/llama-3.3-70b-instruct-fp8-fast | 131,072 | — | — | 5 | Yes | Unknown |
| Cloudflare Workers AI | @cf/meta/llama-4-scout-17b-16e-instruct | — | — | — | 5 | Yes | Unknown |
| NVIDIA NIM (build.nvidia.com) | deepseek-ai/deepseek-r1 | 128,000 | 40 | — | 5 | Yes | Unknown |
You’re seeing 25 of 117 endpoints.
These are the free, lowest-friction ones — enough to get running today. A free account opens the whole matrix plus search, sorting, and filters for capability, provider, and data policy. No card, and the catalog itself stays open data either way.
How to read this
- — means we don’t know, not zero. An unknown quota is recorded as unknown rather than guessed; the provider page shows what we checked and when.
- steps is how many actions stand between you and a working key. Lower is faster to start.
- trains? is provider-level policy on whether your prompts may be trained on. “Unknown” is a real answer and we don’t round it to No.
- free distinguishes a genuine free tier from Trial — credits that run out.
The catalog behind this table is open data under CC BY 4.0 — why that matters.