Compare useful capacity, model access, account requirements, rate limits, and post-allowance behavior. Permanent free tiers stay separate from temporary credits and experiments.
“Free” means little without daily tokens, requests, model access, and reset behavior.
Prototype vs production
Experimental endpoints and rotating models are useful, but clearly labeled as unstable.
Official evidence
Every changing claim links to the provider and carries a verification date.
Clear categories
Different kinds of “free” are not interchangeable
Filter verified offers
Verified offers
Showing 9 of 9; commission never affects ranking.
Free Tier
Hetzner
Verified
Free experimental OpenAI-compatible inference
Hetzner's no-cost experimental OpenAI-compatible inference API now documents Qwen3.8-27B alongside Qwen/Qwen3.6-35B-A3B-FP8. This is a meaningful model-catalog expansion to the already-known Hetzner experiment; availability is not guaranteed and the service should not be used for production.
Value
Free hosted inference during experimental status; two documented vision-capable models with 262,144-token context
Who qualifies
Hetzner customers with access to the Experiments platform who create an inference API token
Expiry
not stated; free while the Inference API remains experimental
Smartoolbox does not currently earn a commission from this offer.
Free Tier
OpenRouter
Verified
Free-model router and zero-cost model variants
OpenRouter publishes a `openrouter/free` router that automatically picks a random free model from those matching the request's capabilities, giving zero-cost inference with one model string.
Value
Free router (`openrouter/free`) auto-selects from 20+ zero-cost models; no API key charges, no credit card; typical free-model limits around 20 requests/minute and 200 requests/day (per OpenRouter docs/community listings).
Who qualifies
Any OpenRouter account holder with an API key; no payment required for free router traffic
Smartoolbox does not currently earn a commission from this offer.
Free Tier
Cloudflare Workers AI
Verified
10,000 Workers AI Neurons free each day
Cloudflare's Workers AI limits page, updated September 17, now quantifies default per-task throughput for the known 10,000-neuron daily free allocation. It lists 300 requests/minute for text generation, 3,000 for text embeddings, and 720 for text-to-image, with lower per-model limits possible for beta or paid-only models.
Value
10,000 neurons/day remains free; default limits include 300 text-generation RPM, 3,000 embedding RPM, and 720 text-to-image RPM
Who qualifies
Cloudflare accounts using Workers AI; the free allocation is available on Workers Free, while models marked paid-only use separate limits
Smartoolbox does not currently earn a commission from this offer.
Free Tier
NVIDIA
Verified
Free hosted NIM APIs for prototyping
NVIDIA Build now lists Moonshot AI's Kimi K3 in its hosted NIM model catalog. NVIDIA states that Build models are free to prototype with and require signing in to generate an API key; hosted use is rate-limited and not positioned as production capacity.
Value
Free hosted Kimi K3 prototyping endpoint; exact model-specific quota not stated
Who qualifies
Developers who create and verify an NVIDIA Build account and generate an API key
Smartoolbox does not currently earn a commission from this offer.
Free Tier
Google Gemini API
Verified
Gemini API free tier for selected models
Google now lists Gemini 3.8 Flash on the Gemini Developer API free tier with free input, output, and context-caching tokens. This is a meaningful model-availability update to the existing generic Gemini API free-tier entry.
Value
Gemini 3.8 Flash input, output, and context caching are free on the limited free tier; numeric quotas are account/model dependent
Who qualifies
Developers and small projects using the Gemini API free tier through Google AI Studio
Smartoolbox does not currently earn a commission from this offer.
Free Tier
GroqCloud
Verified
Free GroqCloud API tier
Groq's current official Free Plan Limits table still includes qwen/qwen3.8-27b, but its daily token cap is now 200,000 rather than the 2 million recorded in the master file. The current limits are 30 requests/minute, 1,000 requests/day, 8,000 tokens/minute, and 200,000 tokens/day.
Value
30 RPM, 1,000 RPD, 8K TPM, and 200K TPD; daily tokens reduced from 2M
Smartoolbox does not currently earn a commission from this offer.
Free Tier
Nous Research
Verified
Nous Portal replaces Union Alpha with Space Bunny Alpha in its eight-model free catalog
The authenticated live Nous Portal catalog still contains eight zero-priced hosted models, but stealth/union-alpha has rotated out and stealth/space-bunny-alpha has appeared. This is a meaningful catalog-composition update while the free-model count remains unchanged.
Value
Eight hosted models at $0 prompt and completion pricing: inclusionai/ling-3.0-flash-fin:free, inclusionai/ling-3.0-flash-sante:free, meituan/longcat-2.0:free, poolside/laguna-s-2.1:free, poolside/laguna-xs-2.1:free, stealth/space-bunny-alpha, stepfun/step-3.7-flash:free, and upstage/solar-pro4:free
Who qualifies
Nous Portal account holders with access to the rotating model catalog
Smartoolbox does not currently earn a commission from this offer.
Free Tier
OpenRouter
Verified
OpenRouter adds Space Bunny Alpha as a free 1M-context model
OpenRouter's official free-model collection now lists Space Bunny Alpha at zero input and output token cost. The stealth preview model supports multimodal input, coding, adjustable reasoning, and a 1M-token context window.
Value
$0/M input and $0/M output tokens with a 1M-token context window
Who qualifies
OpenRouter users with an API key, subject to free-model request caps and provider availability
Expiry
not stated; stealth preview and free-model availability may rotate
Smartoolbox does not currently earn a commission from this offer.
Temporary Promotion
Replicate
Verified
Replicate's 30%-off Wan 3.0 promotion is still live on September 25
Replicate's official pricing page still displays its temporary Wan 3.0 hosted-inference discount on September 25, providing fresh proof that the known promotion remains claimable. The page states 30% off but does not publish the underlying rate or a calendar end date.
Value
30% off Wan 3.0 hosted inference; reconfirmed active on 2026-09-25
Who qualifies
Replicate users running Wan 3.0 while the promotion remains displayed
Expiry
this week; exact calendar deadline not stated (observed 2026-09-25)
Smartoolbox does not currently earn a commission from this offer.
Need a durable comparison?
The maintained guide compares model access, useful capacity, card requirements, data treatment, regions, and what happens when the free allowance ends.