// Models · Free models
For developers who want to call LLMs at zero cost: which makers and platforms give free API access, how much, and whether you need a card – on one page.
20 providers free nowlast verified 2026-10-22official links only · numbers only from official docs
The most generous offers. Start with one of these; switch when you hit the limits.
GlobalAlways freeOpenAINo card
Highest-tier free models: Gemini 2.5 Pro and 3.8 Flash, plus multimodal and speech.
Gemini 3.8 Flash · Gemini 3.6 Flash · Gemini 3.5 Flash-Lite · Gemini 2.5 Pro
Limited per project and per model; exact values are shown only in AI S
ChinaAlways freeOpenAI
Most generous permanent free tier in China: free text, vision, image and video models.
GLM-4.7-Flash · GLM-4-Flash-250414 · GLM-Z1-Flash · GLM-4.6V-Flash
Per-model concurrency limits shown in the console rate-limit page; the
GlobalAlways freeOpenAINo card
Fastest, with clearly published limits: 1,000 requests/day, no card.
openai/gpt-oss-120b · openai/gpt-oss-20b · qwen/qwen3.8-27b · whisper-large-v3-turbo
RPM 30 · RPD 1,000 · TPM 8K
GlobalAlways freeOpenAINo card
One key for a dozen-plus free models with the clearest rules: 50/day, 1,000/day after $10 in credits.
Nemotron 3 Ultra 550B · Nemotron 3.5 Super VL · Gemma 4 31B · MiniMax M3.1 Flash
RPM 20 · RPD 50（累计充值满 10 美元后 1,000）
ChinaAlways freeOpenAI
Free small models for every task in China – text, OCR, embeddings, rerank, image and speech.
XingChenAGI/Xing4.0-29B · tencent/Hunyuan-MT-7B · PaddlePaddle/PaddleOCR-VL-1.5 · BAAI/bge-m3
Free models have fixed rate limits shown on each model page in the con
“not published” means the official docs give no number – the console usually shows it after sign-up.
| Provider | Main free models | Free quota | Rate limits | Card | ID check | OpenAI | Training on data | Commercial use | Official docs · verified |
|---|---|---|---|---|---|---|---|---|---|
| Google AI Studio (Gemini API)Model maker · Global · Always free | Gemini 3.8 FlashGemini 3.6 FlashGemini 3.5 Flash-LiteGemini 2.5 ProGemini 2.5 FlashText / Vision / Speech / audio / Embeddings | Free tier: input and output tokens are free of charge. Gemini 4 Argon, 3.1 Pro Preview, Nano Banana image generation, Veo etc. are not on the free tier. | Limited per project and per model; exact values are shown only in AI Studio, with no fixed public numbers or guarantee. Daily quota resets at midnight Pacific. | No | Not required | ✓https://generativelanguage.googleapis.com/v1beta/openai/ | May be usedYes: free-tier content is used to improve Google products. | Not separately restricted; see the Gemini API Additional Terms.Mainland China and Hong Kong are not on the official availability list. | Gemini API 价格（含免费层） ↗Gemini API 速率限制 ↗可用地区 ↗verified 2026-10-22 |
| Zhipu BigModel (GLM Flash free models)Model maker · China · Always free | GLM-4.7-FlashGLM-4-Flash-250414GLM-Z1-FlashGLM-4.6V-FlashGLM-4.1V-Thinking-FlashText / Vision / Image gen / Video gen | The official price list marks input and output of these 8 models as “free”, with no stated total cap. | Per-model concurrency limits shown in the console rate-limit page; the docs list no fixed numbers for free models. | not published | not published | ✓https://open.bigmodel.cn/api/paas/v4/ | not published | not published | API 定价 ↗速率限制 ↗GLM-4.7-Flash ↗verified 2026-10-22 |
| Groq (Free plan)Platform · Global · Always free | openai/gpt-oss-120bopenai/gpt-oss-20bqwen/qwen3.8-27bwhisper-large-v3-turbocanopylabs/orpheus（TTS）Text / Speech / audio | Usable on the Free plan with per-model limits; a payment method is only needed to upgrade to Developer. | RPM 30 · RPD 1,000 · TPM 8Kgpt-oss-120b/20b, qwen3.8-27b: 30 RPM, 1,000 RPD, 8K TPM, 200K TPD; Whisper 2,000 RPD. Limits are per organisation. | No | not published | ✓https://api.groq.com/openai/v1 | not published | not published | Rate Limits ↗Billing FAQs ↗Supported Models ↗verified 2026-10-22 |
| OpenRouter (:free models)Platform · Global · Always free | Nemotron 3 Ultra 550BNemotron 3.5 Super VLGemma 4 31BMiniMax M3.1 FlashInklingText / Vision | Models whose ID ends in :free cost nothing; on 2026-10-22 the models API lists 18 :free models plus the openrouter/free router. | RPM 20 · RPD 50（累计充值满 10 美元后 1,000）Under $10 purchased all-time: 20 RPM, 50 RPD; $10+ purchased: 1,000 RPD. A negative balance blocks free models too. | No | not published | ✓https://openrouter.ai/api/v1 | Depends on modelDepends on the upstream provider; account settings let you exclude providers that train, separately for free models. | not published | API 限额（免费模型速率） ↗提供方日志与训练政策 ↗模型列表 API ↗verified 2026-10-22 |
| SiliconFlow (free models)Platform · China · Always free | XingChenAGI/Xing4.0-29Btencent/Hunyuan-MT-7BPaddlePaddle/PaddleOCR-VL-1.5BAAI/bge-m3BAAI/bge-reranker-v2-m3Text / Vision / Embeddings / Rerank / Image gen / Speech / audio | Models marked “free” on the pricing page cost nothing (billed at ¥0); paid twins carry a Pro/ prefix. | Free models have fixed rate limits shown on each model page in the console. | not published | not published | ✓https://api.siliconflow.cn/v1 | not published | not published | 模型价格 ↗Rate Limits ↗verified 2026-10-22 |
| Mistral (Free mode)Model maker · Global · Always free | See consoleText | New organisations start in Free mode: create API keys and use the included monthly usage for free. | RPS, tokens/min and tokens/month appear on the Admin Limits page; no fixed public numbers, and the model list is shown in the console. | not published | not published | —https://api.mistral.ai/v1API is similar to OpenAI’s but not officially declared compatible. | Used unless you opt outYes: in Free mode inputs and outputs may be used for training; you can opt out at any time. | not published | Usage and limits ↗Subscriptions ↗数据是否用于训练 ↗verified 2026-10-22 |
| Cloudflare Workers AIPlatform · Global · Always free | @cf/zai-org/glm-4.7-flash@cf/google/gemma-4-26b-a4b-it@cf/nvidia/nemotron-3-120b-a12bEmbedding / 图像 / 语音模型Text / Embeddings / Image gen / Speech / audio | 10,000 Neurons per day on Workers Free (shared across models, resets 00:00 UTC). Since 2026-07-28 a few resource-heavy models need Workers Paid. | Metered in Neurons; token equivalents differ per model (see pricing). | No | not published | ✓https://api.cloudflare.com/client/v4/accounts/{ACCOUNT_ID}/ai/v1 | not published | not published | Workers AI 价格 ↗部分模型需 Workers Paid（2026-07-28） ↗verified 2026-10-22 |
| Alibaba Cloud Model Studio (Qwen)Model maker · China · New-user credit | Qwen 系列（各模型独立额度）Qwen-VLText / Vision / Image gen / Speech / audio | On first activation each model gets its own new-user quota, valid 90 days; Beijing region only. | Real-time inference only (not OpenAI-compatible Batch); unverified users have “stop when free quota runs out” forced on. | not published | Not needed to get or use the free quota; needed for pay-as-you-go afterwards. | ✓https://dashscope.aliyuncs.com/compatible-mode/v1 | not published | not published | 新人免费额度 ↗verified 2026-10-22 |
| Baidu Qianfan (ERNIE)Model maker · China · New-user credit | ERNIE-4.5-Turbo-128KERNIE-4.5-Turbo-VLERNIE-X1-Turbo-32KDeepSeek-V3.1Qwen3-Coder-480BText / Vision / Embeddings | On first activation: 1M tokens for each of 17 listed models, valid 3 months, online inference only. ERNIE Speed / Lite were announced free in Apr 2025 (no longer shown in the current model list – unverified). | Per-model defaults in the model list (e.g. ERNIE 4.5 Turbo: 5,000 RPM, 400K TPM). | not published | not published | ✓https://qianfan.baidubce.com/v2 | not published | not published | 新用户免费额度 ↗模型列表（含默认流控） ↗部分 ERNIE 模型免费开放公告 ↗verified 2026-10-22 |
| Volcengine Ark (Doubao)Model maker · China · New-user credit | 豆包系列等方舟在线推理模型Text / Vision | Registering a Volcengine account grants free inference quota per model, shared across the main account; amounts are shown in the console. | Covers pay-per-token online inference only (not plugins, knowledge bases or batch). “Safe trial mode” stops service before any charge. | not published | Not to claim; needed to continue after the quota runs out. | ✓https://ark.cn-beijing.volces.com/api/v3 | not published | not published | 免费推理额度（更新 2026-08-24） ↗verified 2026-10-22 |
| Tencent Cloud TokenHub (Hunyuan etc.)Model maker · China · New-user credituntil 2026-12-31 | Hy3 preview全部语言模型全部多模态理解模型混元生 3D（积分）Text / Vision / Image gen | Once per main account: 1M tokens for every language and multimodal-understanding model, valid 1 year; offer runs until 2026-12-31. The old hunyuan-lite was retired on 2026-06-22. | If pay-as-you-go is off, service stops when the quota is used up. | not published | not published | ✓https://tokenhub.tencentmaas.com/v1 | not published | not published | TokenHub 新人免费体验包（更新 2026-09-04） ↗混元旧版本模型下线迁移公告 ↗verified 2026-10-22 |
| Cohere (trial key)Model maker · Global · Always free | Command ACommand A ReasoningCommand A VisionCommand R+North Mini CodeText / Vision / Embeddings / Rerank | Trial (evaluation) keys are free, limited to 1,000 API calls per month. | RPM 20Chat 20 req/min per model; Embed 2,000 inputs/min; Rerank 10 req/min; 1,000 calls/month. | not published | not published | ✓https://api.cohere.ai/compatibility/v1 | not published | Trial keys are for evaluation; production needs a paid production key. | API Keys and Rate Limits ↗Cohere 价格 ↗verified 2026-10-22 |
| iFlytek Spark (Lite)Model maker · China · Always free | Spark LiteText | Official docs say Spark Lite “can be used for free”; claim the free quota on the product page. | Max input 8K, max output 4K; daily and concurrency limits per your console authorisation. | not published | not published | ✓https://spark-api-open.xf-yun.com/v1/ | not published | not published | 星火 HTTP 接口文档（兼容 OpenAI SDK） ↗verified 2026-10-22 |
| Kimi Open Platform (Moonshot AI)Model maker · China · New-user credit | Kimi 系列（不含 Kimi K3）Text | ¥15 voucher on sign-up, issued after real-name verification, valid 3 months; not usable for Kimi K3. | not published | not published | Required | ✓https://api.moonshot.cn/v1 | not published | not publishedRequires a mainland China phone number. | 免费试用额度说明 ↗verified 2026-10-22 |
| NVIDIA NIM (build.nvidia.com)Platform · Global · Always free | DeepSeek V4 Pro / FlashKimi K3GLM 5.3 FlashNemotron 系列Text / Vision / Embeddings | Free hosted NIM API endpoints for NVIDIA Developer Program members, for prototyping, research and testing. | Official FAQ: per-model limits that vary with load, shown in your account; no fixed number in the docs. | not published | not published | ✓https://integrate.api.nvidia.com/v1 | not published | Prototyping, research and testing only; production needs an NVIDIA AI Enterprise licence. | Run NIM Anywhere（含价格说明） ↗NVIDIA NIM FAQ ↗verified 2026-10-22 |
| Vercel AI GatewayPlatform · Global · Monthly credit | Free Tier 可用模型（官方列表）Text / Image gen | Every Vercel team gets $5/month of AI Gateway credit, usable on the Free Tier subset of models. | Lower per-model limits on the free tier (no published numbers); BYOK not available on free tier. | not published | not published | ✓https://ai-gateway.vercel.sh/v1 | not published | not published | AI Gateway Pricing ↗verified 2026-10-22 |
| Ollama CloudPlatform · Global · Monthly credit | Starter 模型（云端）Text | Free plan includes starter usage credits for starter models; the amount is not published. | not published | not published | not published | ✓Native API at https://ollama.com/api; OpenAI clients are also supported (subset). | Not usedNo: Ollama states it does not train on cloud prompts and responses. | not published | Ollama 价格 ↗Ollama Cloud 文档 ↗verified 2026-10-22 |
| OpenCode Zen (free models)Platform · Global · Limited-time | GLM-5.3-Flash FreeDeepSeek V4.1 Flash FreeMiniMax M3.1 Flash FreeMiMo-V2.6-Flash FreeNemotron 3.5 Super VL FreeText / Vision | Models whose ID ends in -free are free “for a limited time” and may end any time. | not published | not published | not published | ✓https://opencode.ai/inference/openai/v1 | Depends on modelVaries: GLM-5.3-Flash Free and DeepSeek V4.1 Flash Free providers are zero-retention and do not train; some models state they collect feedback to improve. | not published | OpenCode Zen 文档（免费模型说明） ↗模型列表 API ↗verified 2026-10-22 |
| Cerebras (free trial)Platform · Global · Trial (card) | gpt-oss-120bqwen-3.8-27bText / Vision | One-time $5 credit after adding a verified payment method, valid 30 days; officially no permanent free tier. | RPM 5 · TPM 30KTrial: 5 RPM, 30K uncached TPM, 1M tokens/day. | Required | not published | ✓https://api.cerebras.ai/v1 | not published | not published | Rate Limits（含 Free Trial FAQ） ↗verified 2026-10-22 |
| ModelScope API-InferencePlatform · China · Always free | 魔搭上热门开源模型（Qwen、DeepSeek、GLM 等，标有 API-Inference 图标）Text / Vision / Image gen | Calls are paid in Magicubes: 200/day for logging in, +50/day after linking Alibaba Cloud (valid same day); 0.5 / 1 / 2 per call for light / standard / flagship models – roughly 125–500 calls a day. | Concurrency is limited dynamically by load, guaranteeing single-request use. | not published | Required | ✓https://api-inference.modelscope.cn/v1Requires a linked, ID-verified Alibaba Cloud account. | not published | Officially non-commercial and non-profit; do not use for high-concurrency or SLA-bound production. | API-Inference 使用限制 ↗魔粒说明 ↗verified 2026-10-22 |
No provider matches these filters.
Often recommended online, but checked against official docs they no longer offer a free tier.
Caveats
Method
Data updated: Free APIs(UTC+8)