Completely free, completely free, the most comprehensive collection of free AI APIs on the entire web

Info: Machine translation This post was machine-translated from the Chinese original. Wording may be rough in places — the Chinese version is authoritative.

Note Last retest: 2026-08-23 · HyphenTech　| Request official documentation pages and interfaces from each vendor, with screenshots anchored at the free quota evidence; If you can’t get a hard number, mark it all and check it in the console—no fabrication.

Note This table only includes official free quotas issued to ordinary users. Leaked private keys, shared accounts of unknown origin, quotas obtained through quota transfer stations, fraudulent or forged regions, will not be included, tested, or forwarded.

📋 ** Quick Reference Table **

Note Each of the following points shows: base_url that can be directly copied, call code, features, and the most common restrictions. The screenshots are anchored at the free quota evidence section on each company’s official page.

OpenRouter

One Key connects 22 zero-price models, with the widest range of aggregation entry points among free quotas

Features

• 2026-08-23 Actual Take: Out of 422 models across the site, 22 have dual zero pricing for input and output, with 5 contexts reaching 1 million

• 18 with the ‘free’ suffix (limited by the platform’s free tier), 4 without (vendor-side pricing is 0, not included in this cap)

• Four without extensions: stealth/ox-alpha, google/lyria-3-pro-preview, google/lyria-3-clip-preview, openrouter/free

• NVIDIA alone accounts for 8 out of 22 free models

Note Not being restricted by the platform’s free tier does not mean unlimited traffic: providers may still return 429, Cloudflare’s DDoS passport is always active, and when the account balance is negative, the free model also shows errors.

from openai import OpenAI
client = OpenAI(api_key="<KEY>", base_url="https://openrouter.ai/api/v1")
r = client.chat.completions.create(
    model="nvidia/nemotron-3-ultra-550b-a55b:free",
    messages=[{"role":"user","content":"你好"}])

SiliconFlow

The most generous provider in China for direct connection and free model speed limits; free and paid versions are distinguished by model prefixes

Features

• ⚠️ The most common pitfall: the free version uses the original model name (Qwen/Qwen2.5-7B-Instruct), while the paid version adds the Pro/ prefix (Pro/Qwen/Qwen2.5-7B-Instruct) before the name. If you fill in the wrong prefix, the free version becomes a fee.

• The speed limit for free models is fixed and does not rank according to account usage; Only paid models fluctuate according to usage level

• After calling the free model, the bill shows a fee of 0

Note When RPM, RPH, RPD, TPM, TPD, IPM, IPD reaches peak first, rate limiting is triggered, not just by the number of requests.

from openai import OpenAI
client = OpenAI(api_key="<SILICONFLOW_KEY>", base_url="https://api.siliconflow.cn/v1")
r = client.chat.completions.create(
    model="Qwen/Qwen2.5-7B-Instruct",   # 免费版：无 Pro/ 前缀
    messages=[{"role":"user","content":"你好"}])

BigModel Zhipu

The only domestic platform with a complete set of four items: text, visuals, text-to-image, and text-to-video

Features

• Text: GLM-4.7-Flash, GLM-4-Flash-250414 (128K context / 16K output)

• Vision: GLM-4.6V-Flash, GLM-4.1V-Thinking-Flash, GLM-4V-Flash (64K / 16K)

• Text-to-image image: CogView-3-Flash

• Text-to-Video Video: CogVideoX-Flash

• The same company has gathered all four types of free versions for the same mode, which is rare among free limits

Note Free models will be discontinued with each version iteration (the documentation already marks the model as “going offline”). Before integration, check the current status on the model overview page.

from openai import OpenAI
client = OpenAI(api_key="<ZHIPU_KEY>",
                base_url="https://open.bigmodel.cn/api/paas/v4")
r = client.chat.completions.create(
    model="glm-4-flash-250414",
    messages=[{"role":"user","content":"你好"}])

ModelScope

Register and use immediately, no real-name real-name required for domestic inference entry; both text and text-to-image follow the same OpenAI-compatible interface

Features

• Text-to-image uses v1/images/generations, while asynchronous tasks use v1/tasks/{task_id} for polling

• When higher concurrency is needed, external providers can be bound via API-Provider capabilities

Note The official figure has not been released for a unified quota; high-concurrency scenarios must be confirmed on the console manually.

from openai import OpenAI
client = OpenAI(api_key="<MODELSCOPE_ACCESS_TOKEN>",
                base_url="https://api-inference.modelscope.cn/v1/")

Groq

The lowest latency among the free tiers, but what really lags is TPM, not request count

Features

• GPT-OSS series: 30 RPM / 1K RPD / 8K TPM / 200K TPD

• Orpheus series: 10 RPM / 100 RPD / 1.2K TPM / 3.6K TPD

• Besides text, it can also run Whisper’s speech recognition

Note ⚠️ TPM is the real gate. A single long-context request can eat up an entire minute’s token quota, manifesting as ‘429 after just a few times.’ Batch tasks should be counted as concurrency according to TPM, not by RPM.

from openai import OpenAI
client = OpenAI(api_key="<GROQ_KEY>",
                base_url="https://api.groq.com/openai/v1")

Cloudflare Workers AI

The only free tier based on computing power rather than usage, with the broadest coverage

Features

• Neuron is a unit of computing power, not a number of requests, and many small requests are especially cost-effective

• Covers LLMs, embeddings, text-to-image, speech recognition, speech synthesis, classification, translation, and vision

Note ⚠️ Overrunning is neither slowdown nor automatic charge, but a direct error failure; Afterwards, pricing is $0.011 / 1,000 Neurons. Running a large image or long audio session can consume a large portion of the daily quota at once. Some advanced models require binding payment methods.

// Worker 内直接调用
const r = await env.AI.run('@cf/meta/llama-3.1-8b-instruct', {
  messages: [{ role: 'user', content: '你好' }]
});

Agnes AI

Rare images and videos included in the free quota are provided, but there are three document errors in endpoints and fields

Features

• Image model: agnes-image-2.1-flash, video model: agnes-video-v2.0

• My test on 2026-08-23: image output in 41 seconds, 2624×1472

• Image-to-video is about 100 seconds, pure text-to-video is about 292 seconds — if you can give the first frame, don’t use pure text

Note ⚠️ Four pitfalls you must face: (1) The domain name is apihub.agnes-ai.com, written as API. Directly 404; (2) The video result URL is in the response top-level URL, but the metadata.url written in the document does not exist; (3) Image size only supports 1K/2K/3K/4K levels and ratio; sending 1920x1080 will be standardized to 1312×736; (4) num_frames must be ≤441 and meet 8n+1 requirements. Additionally, local image input must be converted to data URI. Using a public URL can create tasks, but the product is corrupt. The return file extension is .jpg, but the actual encoding may be PNG, so you must truly transcode it before forwarding to WeChat.

# 视频：建任务 → 轮询 → 从响应顶层 url 下载
# POST https://apihub.agnes-ai.com/v1/videos
# GET  https://apihub.agnes-ai.com/agnesapi?video_id=<id>

Hugging Face Inference Providers

The quota is very small, suitable for connectivity testing rather than production

Features

• Quotas are only automatically deducted when routing requests through HF; Direct connections do not apply to all providers

• When your quota runs out, you can purchase additional credits to continue using

Note ⚠️ The $0.10 per free user per month is a very small amount, and the official note is that it may change. When used for connectivity testing and toy projects, don’t use it as a production gateway.

from openai import OpenAI
client = OpenAI(api_key="<HF_TOKEN>",
                base_url="https://router.huggingface.co/v1")

NVIDIA NIM (build.nvidia.com)

A whole row of models labeled as Free Endpoints, perfect for trying out new models

Features

• Many entries on the model plaza are directly marked as Free Endpoint

• NVIDIA also ranks 8 out of 22 on the OpenRouter free list, making it the most aggressive in free deployment

Note The official number of free credit limits is not provided on the public page; you must verify it on the console before connecting to production.

from openai import OpenAI
client = OpenAI(api_key="<NVIDIA_KEY>",
                base_url="https://integrate.api.nvidia.com/v1")

Google Gemini API

The free tier does not require binding a settlement account, but each model’s quota must be checked by yourself in AI Studio

Features

• Rate limits are applied by project, not by API keys

• Exceeding any one of RPM, TPM, RPD, will trigger rate limiting

• The raw image model also has an IPM (Images Per Minute) metric

Note ⚠️ This table does not provide specific limits for Gemini—the official statement states that limits automatically change with user level and account status, and any numbers written in death will expire.

from openai import OpenAI
client = OpenAI(api_key="<GEMINI_KEY>",
    base_url="https://generativelanguage.googleapis.com/v1beta/openai/")

Alibaba Cloud Bailian Model Studio

New users receive 100+ million tokens, but this is a one-time trial fee, not a permanent free tier

Features

• The Tongyi Qianwen system is entirely useful

• Large quota, suitable for running batch tasks all at once

Note ⚠️ This is ONE_TIME_TRIAL, not PERMANENT_FREE_TIER. Calling it ‘forever free’ is the most common lie in these kinds of tutorials. You have to pay once you use it.

from openai import OpenAI
client = OpenAI(api_key="<DASHSCOPE_KEY>",
    base_url="https://dashscope.aliyuncs.com/compatible-mode/v1")

iFlytek Spark

There are permanent free model tiers, as well as free packs that need to be claimed manually

Features

• Lite, Pro, Pro-128K, Max, Max-32K, and 4.0 Ultra each measure tokens independently

• 1 token is approximately equivalent to 1.5 Chinese characters or 0.8 English words

Note ⚠️ Authentication is a three-piece setup of AppID + APIKey + APISecret, and it runs via WebSocket. Unlike other providers, you can’t just switch to a base/_url and connect, which is significantly more expensive to access.

# 非 OpenAI 兼容：需按官方 WebSocket 鉴权流程签名后连接
# 详见 https://www.xfyun.cn/doc/spark/Web.html

Cerebras

Register and receive a $5 trial fee, a one-time bonus, not permanent free

Features

• Renowned for his extremely fast reasoning speed

• The paid tier starts at $10, with a speed limit ten times that of the free tier

Note ⚠️ ONE_TIME_TRIAL. $5 is available while supplies last; don’t plan for permanent free instead.

from openai import OpenAI
client = OpenAI(api_key="<CEREBRAS_KEY>",
                base_url="https://api.cerebras.ai/v1")

Together AI

Dynamic Speed Limiting, No Fixed Free Quota

Features

• You can query your current real-time speed limit through the interface

• Suitable for long-term tasks with stable traffic, not for sudden batch calls

Note ⚠️ There is no fixed free limit number to be cited. Any tutorial claiming ‘how many times a day’ for the Together free tier is questionable; use the real-time speed limit found on your own account as the standard.

from openai import OpenAI
client = OpenAI(api_key="<TOGETHER_KEY>",
                base_url="https://api.together.xyz/v1")

Open Source Hosting (Ollama / llama.cpp / whisper.cpp / Piper)

The path without worrying about others’ opinions: the credit limit is your hardware, not someone else’s policy

Features

• Ollama provides OpenAI-compatible endpoints, allowing you to reuse existing code by switching base_url

• llama.cpp Responsible for reasoning, whisper.cpp Speech recognition, Piper for speech synthesis

• The only solution that won’t fail overnight just because the other party changes its policy

Note The cost is hardware and operations: model size is limited by video memory, speed is limited by local computing power. It solves the problem of “not being cut off,” not “faster and stronger.”

from openai import OpenAI
client = OpenAI(api_key="ollama", base_url="http://localhost:11434/v1")
r = client.chat.completions.create(model="qwen3",
    messages=[{"role":"user","content":"你好"}])

Pollinations

Keyless anonymous layer is still alive, but I tested two results on the same day

Features

• The Models interface entry is marked with “tier”: “anonymous”, and the official notice states that anonymous requests are not affected by the deprecated plan

• OpenAI is compatible with POST/OpenAI paths and also supports direct read GET formats

• ⚠️ My two tests that day had different results: at 12:10 CST, HTTP 402 was returned (budget too low); at 15:57 CST, the same exit IP was tested four times, all of which reached HTTP 200

Note ⚠️ This is the most important methodological sample in this table: on the same day, with the same exit IP, the same request body, it was rejected in the morning and passed in the afternoon. The status of the free entry can change within a few hours; snapshots at any point in time (including this table) cannot be considered long-term conclusions—please send a real request confirmation before calling.

curl -s https://text.pollinations.ai/openai \
  -H 'Content-Type: application/json' \
  -d '{"model":"openai-fast","messages":[{"role":"user","content":"ping"}]}'
# 不带任何 Key；先自测一次再决定是否接入

GitHub Models

2026-07-30 Fully retired, full network tutorials now full of dead links

Features

• The official migration direction is Azure AI Foundry or GitHub Copilot

Note Official original text: As of July 30, 2026, GitHub Models has been fully retired. The playground, model catalog, inference API, and bring your own key (BYOK) are no longer available to any customer.

# 已退役，无调用方式

Note The free list has no expiration date, only the last verification date. This table was retested on 2026-08-23; Please send a real request to confirm the current status before calling it.

🧰 Tools I build

I maintain all of these tools myself. Preview builds are clearly labeled; the release pages are the source of truth for downloads, updates and known limits.

Info: HyphenBox Status: Official releases

A radar for free LLM APIs: availability is re-tested continuously, one local interface for all of them, and keys stay on your machine

Downloads & updates

Info: LocalBrain Status: Official releases

A multimodal MCP toolbox for local models: TTS, Whisper and video generation in one place

Downloads & updates

Info: ScreenLex Status: Official releases

Learn new words while you watch shows. Free, for Mac and Windows

Downloads & updates

Info: HyphenScreen Status: Official releases

Screen recording and smart editing in one: a DaVinci-style timeline, automatic redaction and a check of the finished video before export. Free

Downloads & updates

Quote: HyphenTech Make AI your superpower Local deployment · Free resources · Self-made software https://hyphentech.top