Reference · Free tiers only

Free AI models & providers

Every provider here has a permanent free tier or is free outright — no trial credits, no time-limited promos. Compiled by searching and scraping live provider directories, then separating what is genuinely free from what quietly requires a card, a quota, or identity verification.

scoped to 2026 providers 16 models listed 120+ currency USD compiled 2026-09-15

16Providers

5First-party APIs

11Inference hosts

3No key at all

01 First-party APIs

APIs run by the companies that train or fine-tune the models themselves. All are OpenAI-SDK compatible unless noted, and none require a credit card.

Google Gemini — free tier · US · `generativelanguage.googleapis.com/v1beta`
Model	Context	Max out	Modality	Free limit
Gemini 3.7 Flash	1M	65K	Text · Image · Audio · Video	—
Gemini 3.6 Flash	1M	65K	Text · Image · Audio · Video	15 rpm · 1 500 rpd
Gemini 3.5 Flash-Lite	1M	65K	Text · Image · Audio · Video	30 rpm · 1 500 rpd
Gemini 2.5 Pro	1M	65K	Text · Image · Audio · Video	5 rpm · 50 rpd
Gemma 4 31B	256K	32K	Text	—

Mistral AI — free mode · FR · `api.mistral.ai/v1`
Model	Context	Modality	Free limit
Mistral Large 3	256K	Multimodal	~1 rps · 500K tpm
Mistral Medium 3.5 · 128B	256K	Text · Image · Code	~1 rps · 500K tpm
Mistral Small 4	256K	Text · Image · Code	~1 rps · 500K tpm
Codestral	128K	Code	~1 rps · 500K tpm
Ministral 3 · 3B / 8B / 14B	256K	Text · Vision	~1 rps · 500K tpm

Worth knowing: the Mistral free mode grants roughly $10/month in credits and is on by default — but free-mode prompts may be used to train Mistral models unless you opt out.

Z AI (Zhipu) — permanently free models · CN · `open.bigmodel.cn/api/paas/v4`
Model	Context	Max out	Modality	Free limit
GLM-4.7-Flash	200K	128K	Text · reasoning	1 concurrent
GLM-4.6V-Flash	128K	32K	Multimodal	1 concurrent
GLM-4.5-Flash	128K	96K	Text · reasoning	1 concurrent · retiring

The same free models are served internationally at api.z.ai/api/paas/v4. Registration accepts overseas numbers and the chat API does not require real-name verification.

Cohere — trial key · CA · `api.cohere.com/v2`
Model	Context	Modality	Notes
Command A+ · 218B	128K	Text · Image	20 rpm
Command A · 111B	256K	Text	20 rpm
Command A Reasoning	256K	Text · reasoning	20 rpm
Command A Vision	128K	Text · Image	20 rpm
Aya Expanse 32B · Aya Vision 32B	128K / 16K	Text / Text · Image	20 rpm

non-commercial Cohere's trial key allows 1 000 calls/month for non-commercial use only. No credit card required.

Aion Labs — permanent free tier · IL · `api.aionlabs.ai/v1`
Model	Context	Modality	Free limit
aion-labs/aion-3.0	128K	Text · reasoning	15 rpm · 20K tpd
aion-labs/aion-3.0-mini	128K	Text · reasoning	15 rpm · 20K tpd
aion-labs/aion-rp-llama-3.1-8b	32K	Text · roleplay	15 rpm · 20K tpd

No credit card. Specialised for roleplay and storytelling, with a small daily token budget.

02 Inference hosts

Third-party platforms hosting open-weight models from many sources. This is where the widest free model coverage lives.

Groq — free tier · US · `api.groq.com/openai/v1` · ultra-fast LPU inference
Model	Context	Max out	Free limit
openai/gpt-oss-120b	131K	65K	30 rpm · 1 000 rpd
openai/gpt-oss-20b	131K	65K	30 rpm · 1 000 rpd
qwen/qwen3.6-27b	131K	16K	30 rpm · 1 000 rpd
groq/compound · compound-mini	131K	8K	30 rpm · 250 rpd

NVIDIA NIM — free with Developer Program · US · `integrate.api.nvidia.com/v1` · 100+ models
Model	Context	Max out	Free limit
nvidia/nemotron-3-super-120b-a12b	1M	262K	40 rpm · 10 000 rpd
nvidia/nemotron-3-ultra-550b-a55b	1M	262K	40 rpm · 10 000 rpd
openai/gpt-oss-120b · gpt-oss-20b	131K	131K	40 rpm · 10 000 rpd
mistralai/mistral-nemotron	128K	8K	40 rpm · 10 000 rpd
google/gemma-4-31b-it	262K	8K	40 rpm · 10 000 rpd
+ 92 more	—	—	text · image · video · speech

Cloudflare Workers AI — 10 000 neurons/day · US · `api.cloudflare.com/…/ai/run` · 75+ models
Model	Context	Modality	Free limit
@cf/openai/gpt-oss-120b	128K	Text	shared 10K neurons/day
@cf/meta/llama-4-scout-17b-16e-instruct	131K	Multimodal	shared
@cf/google/gemma-4-26b-a4b-it	256K	Text · Vision	shared
@cf/zai-org/glm-4.7-flash	131K	Text	shared
@cf/deepseek-ai/deepseek-r1-distill-qwen-32b	80K	Text · reasoning	shared
+ 72 more	—	Text · Image · Audio · Embeddings	shared

The daily neuron budget is shared across all usage, not per model, and resets at 00:00 UTC. Exceeding it fails the request rather than billing you. Five models are excluded from the free plan.

OpenRouter — 17 free models · US · `openrouter.ai/api/v1`
Model	Context	Modality	Free limit
nvidia/nemotron-3-super-120b-a12b:free	262K	Text	20 rpm · 50 rpd
openai/gpt-oss-20b:free	131K	Text	20 rpm · 50 rpd
google/gemma-4-31b-it:free	262K	Text · Image	20 rpm · 50 rpd
cohere/north-mini-code:free	256K	Text · code	20 rpm · 50 rpd
poolside/laguna-xs-2.1:free	262K	Text · code	20 rpm · 50 rpd
nvidia/nemotron-nano-12b-v2-vl:free	128K	Text · Image	20 rpm · 50 rpd
+ 11 more free models	—	Text · Image	20 rpm · 50 rpd

Free models allow 50 requests/day each. A one-time $10 credit purchase unlocks 1 000 rpd while keeping the models free. There is also a free-models router and model-fallback chaining.

Ollama Cloud — free tier · US · `ollama.com/api` · OpenAI-compatible at `ollama.com/v1`
Model	Context	Free limit
deepseek-v4-pro · deepseek-v4-flash	1M	session · weekly
kimi-k3	1M	session · weekly
minimax-m3	512K	session · weekly
gpt-oss:120b · gpt-oss:20b	128K · 131K	session · weekly
mistral-large-3:675b	256K	session · weekly
qwen3.5:397b	256K	session · weekly

Usage is measured by weighted input plus output tokens. Session limits reset every 5 hours, weekly limits every 7 days. Exact numbers are not published.

Kilo Code — free pool, no API key · US · `api.kilo.ai/api/gateway` · 200 req/hour
Model	Context	Max out	Modality
nvidia/nemotron-3-ultra-550b-a55b:free	1M	65K	Text
nvidia/nemotron-3.5-lightning:free	1M	65K	Text
stepfun/step-3.7-flash:free	262K	262K	Text · Vision
poolside/laguna-s-2.1:free · laguna-xs-2.1:free	262K	32K	Text · code
tencent/hy3:free	262K	128K	Text
nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free	256K	65K	Multimodal

Reachable with no API key at all. An auto-router picks from the free pool. The catalogue lags what is actually served, so models absent from it have still been observed answering.

Hugging Face — inference credits · US · `router.huggingface.co/v1`
Route	Modality	Free limit
Inference providers — Fireworks · Together · Hyperbolic · Nebius · Novita · DeepInfra	Text · Image · Audio	$0.10/month credits
Meta-Llama-3.1-8B-Instruct	Text · 128K	credit-metered
phi-4 · Qwen2.5-Coder-7B · Qwen2.5-7B	Text	credit-metered
+ thousands of community models	Text · Image · Audio · Embeddings	credit-metered

ModelScope — free inference · CN · `api-inference.modelscope.cn/v1`
Model	Context	Free limit
Qwen/Qwen3.5-35B-A3B	256K	≤500 rpd · 2 000 rpd total
Qwen/Qwen3.5-27B	256K	≤500 rpd · 2 000 rpd total

verification Requires binding an Alibaba Cloud account and passing real-name verification.

LLM7.io — free gateway · GB · `api.llm7.io/v1`
Model	Context	Modality	Anonymous limit
gpt-oss:20b	128K	Text	10 rpm · 60 req/hr
mistral-Nemo-Instruct-2407	128K	Text	10 rpm · 60 req/hr
minimax-m2.7	180K	Text · reasoning	10 rpm · 60 req/hr

A free token (no card) raises limits to 2 rps, 40 rpm, 100 req/hr and 1M tokens/day. The catalogue rotates.

SiliconFlow — permanently free models · CN · `api.siliconflow.cn/v1`
Model	Context	Free limit
Qwen/Qwen3-8B	128K	1 000 rpm · 50 000 tpm

verification Real-name verification is required. Verification supports mainland-Chinese documents; international users must contact support. Most of the 100+ catalogue is paid.

03 No key, no signup

Genuinely zero-friction: no account, no API key, no card. Useful for testing and low-volume work, with tight rate limits.

Provider	Endpoint	Anonymous limit
OVHcloud AI Endpoints · FR	oai.endpoints.kepler.ai.cloud.ovh.net/v1	2 rpm per IP, per model
LLM7.io · GB	api.llm7.io/v1	10 rpm · 60 req/hr
Kilo Code · US	api.kilo.ai/api/gateway	200 req/hr per IP

OVHcloud serves 20+ open-weight models from EU data centres on this anonymous tier, including Qwen3.5-397B-A17B, gpt-oss-120b, Qwen3-Coder-30B-A3B, Qwen2.5-VL-72B, Mistral-Small-3.2-24B and Llama-3.3-70B. Higher limits need a key and are billed per token.

Privacy note Free tiers commonly log prompts. Kilo's own documentation warns its router may send requests to providers that log prompts and outputs, and that NVIDIA's free endpoints are marked trial use — do not submit personal or confidential data. Gemini, Mistral and OpenRouter publish similar caveats.

04 Open-weight models

Weights you can download, inspect and run on your own hardware — no account, no metering, no telemetry. Most of what the free tiers above host originates here.

Family	Latest generation	Strength	Licence
Qwen · Alibaba	Qwen 3.8 · 3.5 series	Best price-to-performance for general inference	Open weights
DeepSeek	DeepSeek V4	Reasoning and coding at frontier-adjacent quality	Open weights
GLM · Zhipu	GLM-5.3 · 4.7-Flash	Strong multilingual and agentic behaviour	Open weights
Kimi · Moonshot	Kimi K3 · K2.6	Long-context and agentic coding	Open weights
gpt-oss · OpenAI	gpt-oss-120b · 20b	Efficient MoE, widely hosted on free tiers	Open weights
Gemma · Google	Gemma 4 · 26B / 31B	Strong small-model quality for local use	Open weights
Llama · Meta	Llama 4 · 3.x	Broad ecosystem and tooling support	Community licence
Mistral · FR	Mistral Large 3 · Small 4	Efficient European models, strong multilingual	Open / research
Nemotron · NVIDIA	Nemotron 3 · Super / Ultra / Nano	MoE families, long context, free on NVIDIA NIM	Open weights
Phi · Microsoft	Phi-4 series	Small models for constrained hardware	Open weights

Practically: open-weight means the weights are free to download — it does not always mean a fully permissive data licence. Check the licence before commercial use, particularly for Llama and some Mistral releases.

05 Image & video generation

Free and open-weight generators you can self-host or reach through the hosts above.

Model	Type	Notes
FLUX.2 Dev	Image · text-to-image	Widely rated the most balanced open image model in 2026
Stable Diffusion 3 / SDXL	Image · text-to-image	The longest-standing open family, huge ecosystem
Qwen-Image	Image · text-to-image	Strong prompt adherence and text rendering
Z-Image	Image	Efficient variant aimed at lower VRAM
Wan	Video	Open video generation

Free hosted image generation also appears inside several providers above — Cloudflare Workers AI serves image and audio models on the shared neuron budget, and NVIDIA NIM's catalogue covers image, video and speech alongside text.

Best route for zero cost For image and video at genuinely no cost, self-hosting open weights is the most durable option: no quota, no logging, no rate limit. The free hosted tiers are better suited to occasional or prototype use.

06 Free chat apps

No API involved — consumer products with permanent free tiers, useful when you need a capable assistant without wiring anything up.

Product	Vendor	Card required
ChatGPT free	OpenAI	no
Claude free	Anthropic	no
Gemini free	Google	no
Perplexity free	Perplexity	no
GitHub Copilot free	GitHub	no
DeepSeek chat	DeepSeek	no
Le Chat free	Mistral AI	no
Qwen Chat	Alibaba	no
Kimi	Moonshot	no

Free consumer tiers are usually rate-limited during peak hours and may apply lower model versions than the paid plan. They are not a substitute for an API when building.

07 Caveats & sources

Read before relying on any free tier

Free tiers change without notice. Models get retired and limits get cut — Groq retired two Llama models in August 2026, and one provider here has a flash model already announced for retirement.
"Free" often means your prompts are training data. Gemini, Mistral, OpenRouter and Kilo all publish this caveat. Do not send confidential data through a free tier.
Verification gates are real. Two providers listed require identity verification — one of them supporting mainland-Chinese documents only.
Regional restrictions apply. The Gemini free tier has region-specific terms around the EEA, Switzerland and the UK.
Catalogue ≠ availability. At least one gateway serves models that do not appear in its own model list, and conversely lists models that no longer answer.

Verification method

Provider lists were gathered by web search and then scraped directly from the source pages using a headless browser pipeline, rather than copied from search summaries. Model rows, context windows and rate limits come from the scraped provider tables. Where a value is genuinely unpublished, this document shows — rather than an estimate.

Sources

github.com/mnfst/awesome-free-llm-apis — permanent free tiers, provider-by-provider tables (primary source)
wotai.co/blog/best-free-llm-apis — live-probed free API directory, 110 models across 16 providers
openrouter.ai/blog/tutorials/free-llm-apis-compared — free-tier comparison
kdnuggets.com · dataiku.com/blog — free LLM API overviews
computingforgeeks.com/open-source-llm-comparison — open-weight model comparison
thundercompute.com · bentoml.com — open image-generation models

Scope Only permanent free offerings are listed. Trial credits, time-limited promotions and free-to-start-but-paid-later products were excluded by design.