The most complete directory of free LLMs, vision models, and embedding models. Compare free tiers, rate limits, and capabilities — then test them instantly in our playground. Stop paying for AI you can get for free.
No payment required. These models offer genuinely free access — some with rate limits, some truly unlimited.
Side-by-side comparison of free tier limits across providers
| Provider | Free models | Rate limit | Best free model | Context | Best for |
|---|---|---|---|---|---|
| Google AI Studio | Gemini Flash, Pro (limited) | 15 RPM, 150K tok/day | Gemini 2.0 Flash | 1M tokens | Long context + vision |
| Groq Cloud | Llama 3.3, Mixtral, Gemma | 30 RPM | Llama 3.3 70B | 128K tokens | Ultra-fast inference |
| Cohere | Command R, Embed v3 | Unlimited (trial) | Command R | 128K tokens | RAG + embeddings |
| Hugging Face | 100+ open models | Serverless, ~1 req/sec | Various | Varies | Experimentation |
| Mistral AI | Mistral Large, Small | 1 req/sec | Mistral Large | 128K tokens | Multilingual + code |
| DeepSeek | V3 (open weights) | Self-host = unlimited | DeepSeek V3 | 128K tokens | Cost + Chinese |
| Meta (Llama) | Llama 3.3 (open weights) | Self-host = unlimited | Llama 3.3 70B | 128K tokens | General purpose |
| ModelHub | Playground (multi-model) | 10/hr, 10/day | Various | Varies | Quick testing |
Practical strategies to cut your AI costs to zero
Build and test your entire application on Google AI Studio's free Gemini Flash tier. Only switch to paid when you need production scale. Most providers offer generous free tiers designed exactly for this.
DeepSeek V3, Llama 3.3, and Qwen 2.5 are free to self-host. A single $0.50/hr GPU instance can serve millions of tokens. For high-volume workloads, self-hosting is 99% cheaper than API calls.
Use a free model (Gemini Flash, Groq Llama) for 80% of simple queries, and only call paid models (Claude, GPT-4o) for complex reasoning. ModelHub's AI Router can automate this split — saving up to 90% on costs.
Embedding models like Cohere Embed v3 are free with unlimited usage. Pre-compute and cache embeddings for your knowledge base. For chat, cache common Q&A pairs to eliminate redundant API calls entirely.
Don't rely on one provider. Combine Google's free Gemini (15 RPM) + Groq's free Llama (30 RPM) + Cohere's free Command R (unlimited) = 45+ RPM across 3 providers, all free. ModelHub's API Gateway can unify them under one endpoint.
For 70% of tasks, a 7B-13B model performs nearly as well as a 70B+ model — at 1/10th the cost. Use our Rankings page to find the smallest model that meets your quality bar. Phi-3 (3.8B) runs on a phone for free.