Free AI Models
Browse free API tiers, open-weight models, no-signup tools and limited trials, ranked by the people who use them. Each row shows how to use it, how it performs, the limits and the catch.
43 free optionsLoading votes…No login, one vote per model per device
| # | Model | Modality | Access | Performance | Free limits | The catch | Votes | How to use |
|---|---|---|---|---|---|---|---|---|
| 1 | Gemini 3.8 Flash and familyGoogle AI Studio · ✓ checked 2026-10 | textcodeaudio | Free API | Varies by model | Free on 3.x Flash, 2.5 and Gemma 4; limits shown in AI Studio only | Prompts train Google models outside the EEA, UK and CH✓ no card · ✓ commercial | ||
| 2 | DeepSeek V4 FlashOpen weights · ✓ checked 2026-10 | textcode | Open weight | Strong | Self-host free · MIT | You run it on your own GPUs✓ no card · ✓ commercial | ||
| 3 | gpt-oss-120b, Qwen3.8 27BGroq · ✓ checked 2026-10 | textcodeaudio | Free API | Varies by model | 30 RPM, 1K requests and 200K tokens a day per model | Only 8K tokens a minute on the chat models✓ no card · ✓ commercial | ||
| 4 | GLM-5.3Open weights · ✓ checked 2026-10 | textcode | Open weight | Frontier | Self-host free · GLM-5.3 License (custom, MIT-style) | You run it on your own GPUs✓ no card · ✓ commercial | ||
| 5 | :free models (17 today)OpenRouter · ✓ checked 2026-10 | textcode | Free API | Varies by model | 20 RPM; 50 a day, or 1,000 after buying $10 of credit | Some upstream providers may train on prompts✓ no card · ? commercial | ||
| 6 | Kimi K2.6Open weights · ✓ checked 2026-10 | textcodeimage | Open weight | Strong | Self-host free · Modified MIT | You run it on your own GPUs✓ no card · ✓ commercial | ||
| 7 | Qwen3.8-27BOpen weights · ✓ checked 2026-10 | textcodeimagevideo | Open weight | Strong | Self-host free · Apache-2.0 | You run it on your own GPUs✓ no card · ✓ commercial | ||
| 8 | Mistral Small and familyMistral · ✓ checked 2026-10 | textcode | Free API | Varies by model | Free mode for new accounts; limits shown in the console only | May train on your data unless you opt out✓ no card · ? commercial | ||
| 9 | gpt-oss-120bOpen weights · ✓ checked 2026-10 | textcode | Open weight | Solid | Self-host free · Apache-2.0 | You run it on your own GPUs✓ no card · ✓ commercial | ||
| 10 | Workers AI modelsCloudflare · ✓ checked 2026-10 | textimageaudioembeddings | Free API | Varies by model | 10,000 Neurons a day, reset at 00:00 UTC | Calls fail until the reset once the day is used? no card · ✓ commercial | ||
| 11 | Tencent Hy3Open weights · ✓ checked 2026-10 | textcode | Open weight | Strong | Self-host free · Apache-2.0 | You run it on your own GPUs✓ no card · ✓ commercial | ||
| 12 | API Catalog free endpointsNVIDIA · ✓ checked 2026-10 | textimageembeddings | Free API | Varies by model | 40 RPM and 10,000 requests a day | Testing only, no production; prompts may train models? no card · ✕ commercial | ||
| 13 | DeepSeek V4 ProOpen weights · ✓ checked 2026-10 | textcode | Open weight | Strong | Self-host free · MIT | You run it on your own GPUs✓ no card · ✓ commercial | ||
| 14 | Gemma 4 31B ITOpen weights · ✓ checked 2026-10 | textcodeimage | Open weight | Solid | Self-host free · Apache-2.0 | You run it on your own GPUs✓ no card · ✓ commercial | ||
| 15 | Llama 4 Scout 17B-16E InstructOpen weights · ✓ checked 2026-10 | textcodeimage | Open weight | Light | Self-host free · Llama 4 Community License | You run it on your own GPUs✓ no card · ✓ commercial |
Need more than a free tier?
We put free and open-weight models into production.
Failover across providers, cost caps and checks on your real inputs. A free 30-minute call.
Browse by provider
Free tier or a cheap paid API?
What does running it on a free tier actually cost?
Free tiers are free to call, not free to run. Put in your numbers: we price the hours a free tier adds and the paid overflow, against paying a cheap model for every request.
Free tier route
$450 /mo
6.0 h × $75 × 1 person
Pay GPT-6 Luna for every request
$1.75 /mo
5,000 requests × $0.0003 each at list price
Cheaper at this volumeThe paid API comes out about $448 a month cheaper.
Token counts per request follow the use case (1,500 in, 400 out). Hours are assumptions you can change; prices are list prices.
Free tier updates
Know when a free tier is added, cut or closed.
We use your email only for these updates.
Questions
Free AI models, honestly answered
Are free AI models really free?
Three kinds of free exist: open-weight models you run yourself (you pay for the hardware), free API tiers (free inside rate limits) and no-signup tools. The catch is usually rate limits, a non-commercial clause or your prompts being used for training. Every row shows its catch.
What is the best free AI model right now?
Sort by best performance for raw quality, measured by the Artificial Analysis Intelligence Index, or by top voted for what people using this page rate highest. For quality per dollar at scale, a strong open-weight model on your own hardware is hard to beat.
Can I use free AI models commercially?
Often yes for open weights under Apache-2.0 or MIT. Several free API tiers are for testing only (NVIDIA’s API catalog, Cohere’s trial keys) or use your prompts for training (Gemini’s free tier outside the EEA, UK and Switzerland). Check the catch column and the licence before you ship.
Do free API keys without signup exist?
A few services work with no key at all: they are marked No signup, and their limits are low. Never use shared keys posted online: they bill someone else and get revoked.
How do I use a free model in my project?
Most free tiers speak the OpenAI API, so you change the base URL and key in the SDK you already use. Each row’s How panel has the steps, a quickstart and a prompt you can paste into Claude Code.
How we rank and verify. Rows are ordered by votes (one per model per device, no login); ties keep our default order. Best performance sorts by the Artificial Analysis Intelligence Index of each open-weight model. Free limits come from each provider’s own docs and terms, and each row shows the month we checked it.
What we never list. Shared or leaked API keys, reverse-engineered endpoints, or trials that take a card and bill after the trial without saying so.