Free models
Models you can use for free
Every model in the snapshot that costs nothing to call, split by how you get it: a hosted API, your own hardware, or a paid plan that includes it. Expiry dates are the vendor's, not ours.
Models you can use for free
Every model in the snapshot that costs nothing to call, split by how you get it: a hosted API, your own hardware, or a paid plan that includes it. Expiry dates are the vendor's, not ours.
75 models · 23 providers · Prices verified: 2026-10-25
Hosted API47
Zero-priced hosted endpoints. Expect per-minute and per-day request caps that scale with your account's lifetime credits.
| Model | Provider | Context | Free until | Listed by |
|---|---|---|---|---|
| mistral-7b-instruct-v0.2-loraHosted on Cloudflare's edge; billed by neural usage, free within the daily allowance. | Cloudflare Workers AI | 15K | — | litellm |
| gemma-2b-it-loraHosted on Cloudflare's edge; billed by neural usage, free within the daily allowance. | Cloudflare Workers AI | 8K | — | litellm |
| llama-2-7b-chat-hf-loraHosted on Cloudflare's edge; billed by neural usage, free within the daily allowance. | Cloudflare Workers AI | 8K | — | litellm |
| gemma-7b-it-loraHosted on Cloudflare's edge; billed by neural usage, free within the daily allowance. | Cloudflare Workers AI | 4K | — | litellm |
| north-mini-code | Cohere | 256K | — | openrouter |
| dots-3-note-preview | Dots Studio | 512K | 2026-12-31 | openrouter |
| GigaChat-2 | GigaChat | 128K | — | litellm |
| GigaChat-2-Max | GigaChat | 128K | — | litellm |
| GigaChat-2-Pro | GigaChat | 128K | — | litellm |
| gemma-4-26b-a4b-it | 262K | — | openrouter | |
| gemma-4-31b-it | 262K | — | openrouter | |
| gemma-3-27b-it | Google Gemini | 131K | — | litellm |
| learnlm-1.5-pro-experimental | Google Gemini | 33K | — | litellm |
| lfm-2.5-2.6b | Liquid AI | 66K | — | openrouter |
| labs-leanstral-1-5-1 | Mistral | 262K | — | litellm |
| nemotron-3-ultra-550b-a55b | NVIDIA | 1000K | — | openrouter |
| nemotron-3.5-lightning | NVIDIA | 1000K | — | openrouter |
| nemotron-3.5-super-vl | NVIDIA | 1000K | — | openrouter |
| nemotron-3-super-120b-a12b | NVIDIA | 262K | — | openrouter |
| nemotron-3-nano-omni-30b-a3b-reasoning | NVIDIA | 256K | — | openrouter |
| nemotron-3.5-content-safety | NVIDIA | 128K | — | openrouter |
| qwen3-coder:480b-cloudRuns on Ollama Cloud, free within its usage allowance. | Ollama Cloud | 262K | — | litellm |
| deepseek-v3.1:671b-cloudRuns on Ollama Cloud, free within its usage allowance. | Ollama Cloud | 164K | — | litellm |
| gpt-oss:120b-cloudRuns on Ollama Cloud, free within its usage allowance. | Ollama Cloud | 131K | — | litellm |
| gpt-oss:20b-cloudRuns on Ollama Cloud, free within its usage allowance. | Ollama Cloud | 131K | — | litellm |
| laguna-s-2.1 | Poolside | 262K | 2026-10-31ending soon | openrouter |
| laguna-xs-2.1 | Poolside | 262K | 2026-10-31ending soon | openrouter |
| Olmo-3-32B-Think | Public AI | 33K | — | litellm |
| Olmo-3-7B-Instruct | Public AI | 33K | — | litellm |
| Olmo-3-7B-Think | Public AI | 33K | — | litellm |
| Qwen-SEA-LION-v4-32B-IT | Public AI | 33K | — | litellm |
| salamandra-7b-instruct-tools-16k | Public AI | 16K | — | litellm |
| ALIA-40b-instruct_Q8_0 | Public AI | 8K | — | litellm |
| apertus-70b-instruct | Public AI | 8K | — | litellm |
| apertus-8b-instruct | Public AI | 8K | — | litellm |
| Gemma-SEA-LION-v4-27B-IT | Public AI | 8K | — | litellm |
| flint-v1.0 | Recursive | 262K | — | openrouter |
| sarvam-m | Sarvam | 8K | — | litellm |
| inkling | Thinking Machines | 1049K | — | openrouter |
| inkling-small | Thinking Machines | 1049K | — | openrouter |
| Ternary-Bonsai-27B | Together AI | 262K | — | litellm |
| kimi-k2-thinking-251104 | Volcengine | 229K | — | litellm |
| glm-4-7-251222 | Volcengine | 205K | — | litellm |
| deepseek-v3-2-251201 | Volcengine | 98K | — | litellm |
| xing4.0-29b-a4b | Xingchen | 262K | — | openrouter |
| glm-4.7-flash | Z.ai | 200K | — | litellm |
| glm-4.5-flash | Z.ai | 128K | — | litellm |
Self-hosted21
The weights are free; the compute is yours. Ollama and Lemonade run on your machine, Cloudflare runs on its edge but still bills the account for CPU time.
| Model | Provider | Context | Free until | Listed by |
|---|---|---|---|---|
| Qwen3-4B-Instruct-2507-GGUF | Lemonade (local) | 262K | — | litellm |
| Qwen3-Coder-30B-A3B-Instruct-GGUF | Lemonade (local) | 262K | — | litellm |
| gpt-oss-120b-mxfp-GGUF | Lemonade (local) | 131K | — | litellm |
| gpt-oss-20b-mxfp4-GGUF | Lemonade (local) | 131K | — | litellm |
| Gemma-3-4b-it-GGUF | Lemonade (local) | 128K | — | litellm |
| mixtral-8x22B-Instruct-v0.1 | Ollama (local) | 66K | — | litellm |
| codegeex4 | Ollama (local) | 33K | — | litellm |
| deepseek-coder-v2-instruct | Ollama (local) | 33K | — | litellm |
| deepseek-coder-v2-lite-instruct | Ollama (local) | 33K | — | litellm |
| internlm2_5-20b-chat | Ollama (local) | 33K | — | litellm |
| mistral-7B-Instruct-v0.2 | Ollama (local) | 33K | — | litellm |
| mixtral-8x7B-Instruct-v0.1 | Ollama (local) | 33K | — | litellm |
| llama3 | Ollama (local) | 8K | — | litellm |
| llama3:70b | Ollama (local) | 8K | — | litellm |
| llama3:8b | Ollama (local) | 8K | — | litellm |
| llama3.1 | Ollama (local) | 8K | — | litellm |
| mistral-7B-Instruct-v0.1 | Ollama (local) | 8K | — | litellm |
| llama2 | Ollama (local) | 4K | — | litellm |
| llama2:13b | Ollama (local) | 4K | — | litellm |
| llama2:70b | Ollama (local) | 4K | — | litellm |
| llama2:7b | Ollama (local) | 4K | — | litellm |
Inside a subscription7
Free inside a plan you already pay for. The request still consumes plan allowance, so this is not a way to avoid the subscription.
| Model | Provider | Context | Free until | Listed by |
|---|---|---|---|---|
| beam:freeNo daily request limit | Command Code | — | — | vendor |
| glyph-cluster:free100 requests/day | Command Code | — | — | vendor |
| laguna-s-2.1-freeWhile capacity lasts | Command Code | — | — | vendor |
| MiniMax M3.1 Flash | Command Code | — | — | vendor |
| Ornith 1.5 122B | Command Code | — | — | vendor |
| Beam Freelimited time | OpenCode | 1049K | — | vendor |
| longcat-2.5-preview-free | OpenCode | 1000K | — | litellm |
Read the numbers like this
- Free is a price, not a promise: the vendor can rate-limit, queue or withdraw the model at any time, and several entries already carry an end date.
- Hosted free endpoints are rate-limited per account, usually by how much credit you have ever bought. They are fine for evaluation and small jobs, not for production volume.
- Self-hosted entries cost nothing per token because you pay for the machine. The model size decides whether that is a laptop or a GPU node.
- Free inside a subscription means the model row is $0 but every request is still metered against your plan's allowance.
Read the numbers like this
- A plan is not a price per token. The same $20 buys 1.5B tokens of a cheap open model or 40M tokens of a frontier model, and several plans grant a different allowance per model.
- Request-based plans are converted with a stated tokens-per-request assumption, because a request count is not a token count. Treat those rows as an estimate and check your own task size.
- Credits are converted with the vendor's own formula (credits = weighted tokens / divisor, peak/off-peak multipliers applied), so the number is exact for that model but moves with time-of-day.
- Claude, ChatGPT, Cursor, Kimi and MiniMax publish no convertible allowance at all. They stay in the table, sorted last, with their price — an honest gap beats an invented number.
FAQ
Why do some rows show a token count and others say not quantified?
Because plans do not sell tokens. Some sell a dollar allowance (OpenCode, Command Code, OpenDesign), some a token volume (Synthetic, Cerebras, Baidu), some vendor credits with per-model multipliers (Z.ai, Volcengine Agent Plan), some a request count (Volcengine, Alibaba), and the biggest names sell an unquantified "usage" (Claude, ChatGPT, Cursor, Kimi, MiniMax). The Allowance column says which, so an inferred number is never mistaken for a published one.
Why not just compare the price per million tokens?
Input, cached read, cached write and output are priced separately, and their ratios differ by an order of magnitude between providers. A single blended figure at stated assumptions is the only number that compares plans honestly.
I only care about one model. What do I do?
Pick the model family first (DeepSeek, GLM, Claude, Qwen...). The page then lists every vendor that resells that family, what each plan costs, how many tokens the plan buys at your cache hit rate, and the resulting blended unit price.
How accurate is the token estimate?
Request counts are the one estimate on the page: they are multiplied by a stated tokens-per-request figure that you can adjust. Every other conversion uses a number the vendor publishes, and every row links to its source page so you can check it.
Where does this data come from?
Unit prices come from LiteLLM and OpenRouter; subscription prices and allowances are parsed from each vendor's own plan page (or declared with an anchor guard when the page only renders prices in the browser). The verified date is on every row — prices move, so re-check before committing to an annual plan.