本文最后更新于 2026年9月4日。
A reference table of free LLM API providers, verified September 2026. Note that the free-tier landscape has tightened considerably since mid-2026: GitHub Models was fully retired on 2026-07-30, Cerebras removed its permanent free tier, and OpenRouter cut :free quotas from 200 RPD to 50 RPD. Always confirm limits against the x-ratelimit-* response headers rather than the numbers below.
Provider APIs
APIs run by the companies that train or fine-tune the models themselves.
Cohere 🇨🇦
Free "Trial" API key, no credit card. 1,000 API calls/month. Non-commercial use only.
Base URL: https://api.cohere.com/v2
| Model Name | Context | Max Output | Modality | Rate Limit |
|---|---|---|---|---|
| Command A (111B) | 256K | 4K | Text | 20 RPM |
| Command R+ | 128K | 4K | Text | 20 RPM |
| Command R | 128K | 4K | Text | 20 RPM |
| Command R7B | 128K | 4K | Text | 20 RPM |
| Embed 4 | — | — | Embeddings (Text + Image) | 2,000 inputs/min |
| Rerank 3.5 | — | — | Reranking | 10 RPM |
Google Gemini 🇺🇸
Free tier unavailable in EU/UK/Switzerland. Free-tier prompts may be used by Google to improve products. [^1] Image models are not part of this free tier — Gemini 3.x image models require the pay-as-you-go Blaze plan regardless of provider.
Base URL: https://generativelanguage.googleapis.com/v1beta
| Model Name | Context | Max Output | Modality | Rate Limit |
|---|---|---|---|---|
| Gemini 2.5 Pro | 1M | 65K | Text + Image + Audio + Video | 5 RPM, 100 RPD |
| Gemini 2.5 Flash | 1M | 65K | Text + Image + Audio + Video | 10 RPM, 250 RPD |
| Gemini 2.5 Flash-Lite | 1M | 65K | Text + Image + Audio + Video | 15 RPM, 1,000 RPD |
| Gemini 3.1 Flash-Lite | 1M | 65K | Text + Image | 5 RPM, ~500 RPD |
| Gemma 4 26B | 256K | 8K | Text | ~15 RPM, ~1,500 RPD |
Mistral AI 🇫🇷
Free "Experiment" plan, no credit card. ~1B tokens/month.
Base URL: https://api.mistral.ai/v1
| Model Name | Context | Max Output | Modality | Rate Limit |
|---|---|---|---|---|
| Mistral Small 4 | 256K | 256K | Text + Image + Code | ~60 RPM, 500K TPM |
| Mistral Medium 3 | 128K | 128K | Text | ~60 RPM, 500K TPM |
| Mistral Large 3 | 256K | 256K | Text | ~60 RPM, 500K TPM |
| Mistral Nemo (12B) | 128K | 128K | Text | ~60 RPM, 500K TPM |
| Codestral | 256K | 256K | Code | ~60 RPM, 500K TPM |
| Pixtral Large | 128K | 128K | Text + Image | ~60 RPM, 500K TPM |
Z AI (Zhipu AI) 🇨🇳
Permanent free models, no credit card required.
Base URL: https://open.bigmodel.cn/api/paas/v4
| Model Name | Context | Max Output | Modality | Rate Limit |
|---|---|---|---|---|
| GLM-4.7-Flash | 200K | 128K | Text | 1 concurrent request |
| GLM-4.5-Flash | 128K | ~8K | Text | 1 concurrent request |
| GLM-4.6V-Flash | 128K | ~4K | Text + Image | 1 concurrent request |
Inference providers
Third-party platforms that host open-weight models from various sources.
Cerebras 🇺🇸
No permanent free tier — official FAQ answers "Is there a permanently free tier?" with a plain No. Only a Free Trial remains: $5 credits, 30-day expiry, payment method required. Ultra-fast inference (~2,600 tok/s). Rate limits are now split into uncached/total buckets.
Base URL: https://api.cerebras.ai/v1
| Model Name | Context | Modality | Rate Limit (free) |
|---|---|---|---|
| gpt-oss-120b | 128K | Text | 5 RPM, 30K TPM, 1M TPD |
| gemma-4-31b | 128K | Text | 15 RPM, 30K TPM, 1M TPD |
Cloudflare Workers AI 🇺🇸
10,000 Neurons/day free, resets 00:00 UTC, available on both Free and Paid plans. Two caveats: kimi-k2.6, kimi-k2.7-code, glm-5.2/5.3/5.3-flash and deepseek-v4-flash/pro require a bound payment method; and this platform also serves free text-to-image models (@cf/black-forest-labs/flux-1-schnell), which are not subject to that restriction.
Base URL: https://api.cloudflare.com/client/v4/accounts/{account_id}/ai/run
| Model Name | Context | Modality | Rate Limit |
|---|---|---|---|
| @cf/meta/llama-3.3-70b-instruct-fp8-fast | 131K | Text | 10K neurons/day (shared) |
| @cf/meta/llama-3.1-8b-instruct-fp8-fast | 131K | Text | 10K neurons/day (shared) |
| @cf/meta/llama-4-scout-17b-16e-instruct | Up to 10M | Multimodal | 10K neurons/day (shared) |
| @cf/qwen/qwen3.8-27b | 131K | Text | 10K neurons/day (shared) |
| @cf/zai-org/glm-4.7-flash | 128K | Text | 10K neurons/day (shared) |
| @cf/nvidia/nemotron-3-120b-a12b | 262K | Text | 10K neurons/day (shared) |
| @cf/google/gemma-4-26b-a4b-it | 256K | Text | 10K neurons/day (shared) |
| @cf/mistralai/mistral-small-3.1-24b-instruct | 128K | Text | 10K neurons/day (shared) |
| + 42 more models | Varies | Text, Image, Audio, Embeddings | 10K neurons/day (shared) |
Groq 🇺🇸
Free Developer Plan, no credit card. Ultra-fast LPU inference. Figures below are reported consistently by third-party directories but could not be confirmed against Groq’s own console (returned 403). Verify with response headers. [^2]
Base URL: https://api.groq.com/openai/v1
| Model Name | Context | Max Output | Modality | Rate Limit |
|---|---|---|---|---|
| llama-3.3-70b-versatile | 131K | 32K | Text | 30 RPM, 1,000 RPD, 6,000 TPM |
| llama-3.1-8b-instant | 131K | 131K | Text | 30 RPM, 1,000 RPD, 6,000 TPM |
| llama-4-scout-17b-16e-instruct | 131K | 8K | Text + Vision | 30 RPM, 1,000 RPD, 6,000 TPM |
| kimi-k2-instruct | 262K | 262K | Text | 30 RPM, 1,000 RPD, 6,000 TPM |
| whisper-large-v3 | — | — | Audio → Text | 20 RPM, 2,000 RPD |
qwen3-32b and gpt-oss-120b appear to have been removed from the free catalogue.
Hugging Face 🇺🇸
Free Serverless Inference API + ~$0.10/month free credits (PRO: $2.00/month). No free credits when using a Custom Provider Key. At roughly $0.0012 per FLUX.1-dev image, $0.10 covers only about 80 images — fine for validation, not for production.
Base URL: https://router.huggingface.co/v1
| Model Name | Context | Modality | Rate Limit |
|---|---|---|---|
| Meta-Llama-3.1-8B-Instruct | 128K | Text | ~1,000 RPD |
| Mistral-7B-Instruct-v0.3 | 32K | Text | ~1,000 RPD |
| Qwen2.5-7B-Instruct | 131K | Text | ~1,000 RPD |
| + thousands of community models | Varies | Text, Image, Audio, Embeddings | ~$0.10/month free credits |
Kilo Code 🇺🇸
Free models with no credit card required. kilo-auto/free auto-router routes to minimax/minimax-m2.5:free (80%) and stepfun/step-3.5-flash:free (20%). [^5]
Base URL: https://api.kilo.ai/api/gateway
| Model Name | Context | Modality | Rate Limit |
|---|---|---|---|
| nvidia/nemotron-3-super-120b-a12b:free | 262K | Text | ~200 req/hr |
| arcee-ai/trinity-large-thinking:free | — | Text (reasoning) | ~200 req/hr |
| bytedance-seed/dola-seed-2.0-pro:free | — | Text | ~200 req/hr |
| x-ai/grok-code-fast-1:optimized:free | — | Text (code) | ~200 req/hr |
| openrouter/free | Varies | Text | ~200 req/hr |
LLM7.io 🇬🇧
Zero-friction API gateway. No registration needed for basic access. 30+ models.
Base URL: https://api.llm7.io/v1
| Model Name | Modality | Rate Limit |
|---|---|---|
| deepseek-r1-0528 | Text (reasoning) | 30 RPM (120 with token) |
| deepseek-v3-0324 | Text | 30 RPM (120 with token) |
| gemini-2.5-flash-lite | Text + Vision | 30 RPM (120 with token) |
| gpt-4o-mini | Text + Vision | 30 RPM (120 with token) |
| qwen2.5-coder-32b | Text (code) | 30 RPM (120 with token) |
| + ~24 more models | Text | 30 RPM (120 with token) |
ModelScope 🇨🇳
Free API-Inference for registered users. Requires Alibaba Cloud account binding + real-name verification. Currently the largest free quota in this table. [^6]
Base URL: https://api-inference.modelscope.cn/v1
| Model Name | Modality | Rate Limit |
|---|---|---|
| Qwen/Qwen3.5-35B-A3B | Text + Vision | 2,000 RPD total; <=500 RPD/model (dynamic) |
| Qwen/Qwen3.5-27B | Text | 2,000 RPD total; <=500 RPD/model (dynamic) |
| Qwen/Qwen-Image | Image Generation | 2,000 RPD total; model/AIGC-specific caps |
| + API-Inference-enabled models | LLM, MLLM, AIGC | Dynamic quotas + dynamic concurrency |
NVIDIA NIM 🇺🇸
Free with NVIDIA Developer Program membership (free credits, generally no credit card). 100+ models, about 38 of them carry a Free Endpoint. Note that some entries such as QWEN-IMAGE are marked Downloadable (weights only), not free endpoints — filter the catalogue by "Free Endpoint" before relying on one.
Base URL: https://integrate.api.nvidia.com/v1 · OpenAI-compatible: https://ai.api.nvidia.com/v1/
| Model Name | Context | Modality | Rate Limit |
|---|---|---|---|
| deepseek-ai/deepseek-r1 | 128K | Text (reasoning) | ~40 RPM |
| nvidia/nemotron-3-super-120b-a12b | 262K | Text | ~40 RPM |
| nvidia/nemotron-3-nano-30b-a3b | 128K | Text | ~40 RPM |
| meta/llama-3.1-405b-instruct | 128K | Text | ~40 RPM |
| qwen/qwen2.5-72b-instruct | 128K | Text | ~40 RPM |
| google/gemma-4-31b | 128K | Text | ~40 RPM |
| nvidia/nemotron-nano-2-vl | 128K | Vision + Text + Video | ~40 RPM |
| + 90 more models | Varies | Text, Image, Video, Speech, Embeddings | ~40 RPM |
Ollama Cloud 🇺🇸
Free tier with qualitative ("Light") usage limits. Not OpenAI SDK-compatible by default; an OpenAI-compatible endpoint is available at https://ollama.com/v1. [^3]
| Model Name | Context | Modality | Rate Limit |
|---|---|---|---|
| glm-4.7:cloud | 128K | Text | Qualitative / unpublished |
| deepseek-v3.2 | 128K | Text | Qualitative / unpublished |
| qwen3.5 | 128K | Text | Qualitative / unpublished |
| kimi-k2.5 | 256K | Text | Qualitative / unpublished |
| llama3.2 | 128K | Text | Qualitative / unpublished |
OpenRouter 🇺🇸
About 29 free models (marked with the :free suffix). OpenAI SDK-compatible. Quota was cut to 50 RPD; cumulative top-ups of $10 or more raise it to 1,000 RPD while RPM stays at 20. [^4] There are no :free image models — image generation on OpenRouter is a separate, paid API.
Base URL: https://openrouter.ai/api/v1 · Images: https://openrouter.ai/api/v1/images
| Model Name | Context | Modality | Rate Limit |
|---|---|---|---|
| deepseek/deepseek-r1-0528:free | 163K | Text (reasoning) | 20 RPM, 50 RPD |
| deepseek/deepseek-chat-v3-0324:free | 163K | Text | 20 RPM, 50 RPD |
| qwen/qwen3.6-plus:free | 1M | Text | 20 RPM, 50 RPD |
| qwen/qwen3-coder-480b-a35b:free | 262K | Text | 20 RPM, 50 RPD |
| meta-llama/llama-4-scout:free | 10M | Multimodal | 20 RPM, 50 RPD |
| nvidia/nemotron-3-super-120b-a12b:free | 1M | Text | 20 RPM, 50 RPD |
| openai/gpt-oss-120b:free | 131K | Text | 20 RPM, 50 RPD |
| minimax/minimax-m2.5:free | 196K | Text | 20 RPM, 50 RPD |
| mistralai/devstral-2512:free | 256K | Text | 20 RPM, 50 RPD |
| + ~20 more free models | Varies | Text | 20 RPM, 50 RPD |
OVHcloud AI Endpoints 🇫🇷
Free anonymous tier (no API key, no signup): 2 RPM per IP per model. 40+ open-weight models hosted in EU. OpenAI SDK-compatible. [^7]
Base URL: https://oai.endpoints.kepler.ai.cloud.ovh.net/v1
| Model Name | Context | Modality | Rate Limit |
|---|---|---|---|
| Meta-Llama-3_3-70B-Instruct | 131K | Text | 2 RPM (anonymous) |
| DeepSeek-R1-Distill-Llama-70B | 131K | Text (reasoning) | 2 RPM (anonymous) |
| Qwen3-32B | 131K | Text | 2 RPM (anonymous) |
| Qwen3-Coder-30B-A3B-Instruct | 262K | Text (code) | 2 RPM (anonymous) |
| Qwen3Guard-Gen-8B | 32K | Text (safety guard) | 2 RPM (anonymous) |
| + 30 more models | Varies | Text, Vision, Code, Speech | 2 RPM (anonymous) |
SiliconFlow 🇨🇳
Free tier with 14 CNY signup credits (usable for both text and image). Permanently free small models available.
Base URL: https://api.siliconflow.cn/v1
| Model Name | Context | Modality | Rate Limit |
|---|---|---|---|
| Qwen/Qwen3-8B | 131K | Text | 1,000 RPM, 50K TPM |
| deepseek-ai/DeepSeek-R1-0528-Qwen3-8B | ~33K | Text (reasoning) | 1,000 RPM, 50K TPM |
| THUDM/glm-4-9b-chat | 32K | Text | 1,000 RPM, 50K TPM |
| THUDM/GLM-4.1V-9B-Thinking | 66K | Vision + Text | 1,000 RPM, 50K TPM |
| deepseek-ai/DeepSeek-OCR | — | Vision (OCR) | 1,000 RPM, 50K TPM |
Newer gateways (appeared H2 2026)
| Platform | Base URL | Free models / quota |
|---|---|---|
| OpenCode Zen | https://opencode.ai/zen/v1 (OpenAI-compatible) |
minimax-m2.5-free, qwen3.6-plus-free, nemotron-3-super-free, big-pickle; officially described as Unlimited during feedback collection, throttled under congestion |
| routing.run | — | Lite plan 400 req/day, free tier 50 RPD; route/kimi-k2.5, route/glm-5, route/qwen3.5-397b-a17b, route/deepseek-v3.2 |
| Crof | — | qwen3.5-9b, glm-4.7-flash; claims no request cap, 128K context |
| Nous Portal | https://gateway.nous.chat/v1 |
$10 one-time top-up converts to $10/month credits; backed by OpenRouter |
| Fireworks AI | — | $6 credits for new users, 128K context |
Retired
GitHub Models (formerly models.inference.ai.azure.com) was fully retired on 2026-07-30. The playground, model catalogue, inference API and BYOK were all shut down; Microsoft directs users to Azure AI Foundry. Azure AI Foundry itself runs on a separate $200 credit model, which is not usable in GitHub Codespaces.
Glossary
| Abbreviation | Meaning |
|---|---|
| RPM | Requests per minute |
| RPD | Requests per day |
| TPM | Tokens per minute |
| TPD | Tokens per day |
| RPS | Requests per second |
Verification note: the Cloudflare, Cerebras, OpenRouter and GitHub entries come from official first-party pages. Figures for Gemini, Groq, NVIDIA NIM, SiliconFlow and ModelScope could not be confirmed first-hand (official pages timed out or returned 403) and were taken from the third-party directory freetiermodels.com — treat them as directional and confirm with response headers.
相关阅读
- 哪里有免费生图的大模型api——生图方向的专项整理。本表只覆盖文本，出图方案在那边，主力是 Cloudflare Workers AI 与 ModelScope
Qwen/Qwen-Image。 - 免费大模型API汇总与分析——中文长文版，含轮询架构、权重分配与熔断降级方案，适合把本表这些 provider 真正拼成一个可用的网关。
顺手推荐：威聚 Weihub 开源项目库
hub.weiyoun.com（威聚 Weihub）是一个中文开源项目库，收录了 220 个项目，按功能（AI 工具、Web 框架、效率工具、运维、网络工具、数据资源等）和编程语言（Go、Python、Rust、TypeScript、C++ 等）分类，支持按星级、最近更新、名称排序；每个项目带一句 AI 摘要和一个「二开难度」评级（易二开 / 中等 / 难二开）。
对本表读者的用法：这十几个 provider，最后总要落到「一个网关 + 一套 Key 轮换」上。动手之前去 AI 工具和网络工具两个分类按星级翻一遍，经常已经有现成的开源网关可以改——比如项目库里的 Gateway（TypeScript 写的 LLM API 聚合网关，统一访问 200+ LLM，自带多模态、自动重试与负载均衡），正好对得上本表的用途。配合「二开难度」标签筛选，能省掉不少自己造轮子的时间。