Third-party free AI model hosts

⚠ Read this before you use them

• These are resellers and hosts, not the companies that make the models.
• Your prompts and code pass through their servers.
• Don't send private or client code, passwords, API keys or other secrets.
• Free capacity can queue, change or end without notice.
• Read each host's terms and privacy policy before you sign up. We quote the key lines below.
• Quotadex isn't affiliated with any of these hosts and earns nothing from these links.

Free models: qwen/qwen3.8-27b (Qwen3.8-27B) · Only Qwen3.8-27B is free. IteraCompute says the free Qwen3.8-27B promotion ends at 00:00 Pacific Time on November 24, 2026. The other models, including Kimi K3, are paid. source

Limits: 30M tokens/month (input + output, incl. reasoning), 160,000 tokens/min, 128K context, output capped at 8,192 tokens. Resets on the 1st at 00:00 UTC, no rollover. Free accounts may queue at busy times; no uptime SLA. source

Card: No card needed source

Your prompts and code: No training without your explicit permission. Retention of prompts is not stated beyond 'only for the period necessary'. Zero data retention only under a separate written agreement. source

“Without your explicit authorization, we will not use your Model API call data, Interaction Data or Usage Data to train, fine-tune or optimize general-purpose foundation models offered to unspecified users.”
“We provide zero data retention (ZDR) guarantees only under an additional written agreement”

Company: IteraCompute Artificial Intelligence Technology (Chongqing) Co., Ltd., registered in Chongqing, China source

Coding tools (per its own docs): OpenCode, Cline, Roo Code · Setup guides for OpenCode and Cline on its own site; it also links a Roo Code guide. source

base_url = https://api.iteracompute.com/v1

Official free page · Re-check: Pricing page says 'Catalog updated: 2026-10-27' and the free Qwen3.8-27B promotion ends Nov 24, 2026, 00:00 PT. Remove the row after that unless extended.

Free models: glm-5.3-flash-free (GLM 5.3 Flash), deepseek-v4.1-flash-free (DeepSeek V4.1 Flash), gemini-3.8-flash-free (Gemini 3.8 Flash), minimax-m3.1-flash-free (MiniMax M3.1 Flash) · Only the IDs ending in -free are free. The same model without -free is billed. IDs are case-sensitive (the free MiniMax ID is all lowercase). BeatAPI also lists jev-1.14-free on its Decisions API, which is not a chat model. source

Limits: 1 successful request per minute per account before your first top-up, up to 10 per minute after any top-up. No published context or output limit for the free IDs, and no promised daily allowance. glm-5.3-flash-free and minimax-m3.1-flash-free don't support /v1/responses; minimax-m3.1-flash-free can take tens of seconds to return its first token. source

Card: No card or top-up needed source

Your prompts and code: Prompts and outputs may be sent to upstream AI providers. BeatAPI says it doesn't train general-purpose models on private content without notice, but its policy doesn't promise zero retention or no training by every provider. Retention: 'as long as needed'. source

“We do not use private customer content to train general-purpose AI models unless we provide a separate, explicit notice and obtain any consent required by law.”
“This policy does not promise that every provider offers zero retention or excludes all use of data for model training.”

Company: Not stated (no legal entity or country in its Terms or Privacy Policy; contact support@beatapi.io) source

Coding tools (per its own docs): Its OpenCode page covers BeatAPI's MCP data tools, not using the free models as your coding model. No official coding-tool setup for the free models found. source

base_url = https://api.beatapi.io/v1

Official free page · Re-check: BeatAPI says free models 'can be rate limited or withdrawn at any time' and not to build a production path on one. Re-check every few days.

Free models: recursive/flint-v1.0:free, xingchen-agi/xing4.0-29b-a4b:free, nvidia/nemotron-3.5-super-vl:free, dots-studio/dots-3-note-preview:free, liquid/lfm-2.5-2.6b:free, nvidia/nemotron-3.5-lightning:free, thinkingmachines/inkling-small:free, poolside/laguna-s-2.1:free, thinkingmachines/inkling:free, poolside/laguna-xs-2.1:free, cohere/north-mini-code:free, nvidia/nemotron-3.5-content-safety:free, nvidia/nemotron-3-ultra-550b-a55b:free, nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free, google/gemma-4-26b-a4b-it:free, google/gemma-4-31b-it:free · 16 model IDs ending in :free in OpenRouter's public model list on 2026-10-28, including Recursive's Flint coding model (recursive/flint-v1.0:free, 262K context). Nemotron 3 Super's free ID is gone (the paid ID is $0.30/$2.40 per 1M). Poolside's two Laguna :free IDs list an end date of Oct 31, 2026. source

Limits: 20 requests/minute. 50 requests/day if you have bought fewer than 10 credits, 1,000/day after buying at least 10. A negative balance can block free models too. source

Card: No purchase needed for the 50/day tier (card not stated) source

Your prompts and code: OpenRouter itself doesn't store prompts unless you opt in. But each provider behind a model has its own policy, and some may train on prompts. You can block those in Privacy settings, with a separate setting for free models. source

“OpenRouter does not store your prompts or responses, unless you opt in”
“We document which providers may capture data for training, and provide settings to control how your own prompts may be used.”

Company: OpenRouter, Inc., New York, NY, United States source

Coding tools (per its own docs): Cline, Aider, Roo Code, Kilo Code · Listed on OpenRouter's integrations page. source

base_url = https://openrouter.ai/api/v1

Official free page · Re-check: The :free list changes often. Re-pull https://openrouter.ai/api/v1/models and keep only IDs ending in :free.

Free models: openai/gpt-oss-120b, openai/gpt-oss-20b, qwen/qwen3.8-27b · Examples. Groq has a Free tier, and its rate-limit page lists per-model limits for these and other models. source

Limits: Groq's table lists, for openai/gpt-oss-120b and qwen/qwen3.8-27b: 30 requests/min, 1K requests/day, 8K tokens/min, 200K tokens/day. The page also says your exact limits are on the Limits page of your account, and its wording about whether this table is the Free or Developer baseline is unclear. source

Card: No card for the Free tier (a payment method is asked only to upgrade to Developer) source

Your prompts and code: Not retained by default. Inputs and outputs may be logged for up to 30 days for reliability or abuse checks; you can turn on Zero Data Retention. Data is stored in the US. source

“By default, Groq does not retain customer data for inference requests.”
“These logs are retained for up to 30 days, unless legally required to retain longer.”

Company: Groq LLC, Mountain View, California, United States source

Coding tools (per its own docs): OpenCode, Cline, Roo Code, Kilo Code, Factory Droid · Coding guides in Groq's docs. source

base_url = https://api.groq.com/openai/v1

Official free page · Re-check: Re-check the rate-limit table monthly. The page wording on which plan the table is for is unclear; the console Limits page shows your exact numbers.

Free models: beam (Reflection Beam, free for a limited time) · Only Beam is free right now, per Ollama's Oct 16 blog post (ollama.com/blog/reflection-beam) and the 'free' tag in Ollama's cloud model list. MiniMax M3.1 Flash's free period ended Oct 22; it's now billed. Other Ollama cloud models are billed per token from usage credits. source

Limits: Ollama's $0 pay-as-you-go plan has no subscription and allows 3 concurrent requests; other cloud models are billed per token. Beam has a 1M-token context and isn't in Ollama's per-token price list. Ollama's Beam post says it's free for a limited time but gives no end date or token cap, and doesn't say whether Beam needs bought credits. source

Card: The pay-as-you-go plan is $0 with no subscription. Ollama doesn't say whether a card is needed. source

Your prompts and code: Not used for training. Cloud prompts and responses are processed transiently and not stored beyond the request, per Ollama's privacy policy. Ollama's Beam post says its Beam endpoint runs with full zero data retention (ZDR). source

“When using cloud-hosted models, we process your prompts and responses transiently to provide the service and never train on it.”
“this content is not stored beyond the time required to fulfill the request.”

Company: Ollama Inc. (Terms governed by California law, United States) source

Coding tools (per its own docs): Claude Code, OpenCode, Codex CLI · Built-in launch commands in the Ollama app and CLI. source

base_url = https://ollama.com/v1 (OpenAI-compatible, needs OLLAMA_API_KEY)

Official free page · Re-check: MiniMax M3.1 Flash's free window ended Oct 22, 2026 11:59 PM UTC (its page now shows $0.15/$0.60 per 1M). Beam is still 'free for a limited time' per Ollama's blog, with no end date; re-check every few days and end this row when Beam is no longer free.

Free models: deepseek-ai/deepseek-v4.1-flash (DeepSeek V4.1 Flash), z-ai/glm-5-3-flash (GLM-5.3 Flash), nvidia/nemotron-3-ultra-550b-a55b (Nemotron 3 Ultra), mistralai/mistral-large-4-1t-a52b (Mistral Large 4, 'Le Chonk') · Each of these model pages says 'Start building with a free API endpoint.' NVIDIA's catalog has more models with a Free Endpoint label; we list only the coding-relevant ones we checked. source

Limits: Up to 40 requests/min and 10,000 requests/day, per NVIDIA's model pages, which add that limits may vary by model and that traffic from other users may cause throttling. Trial use only: NVIDIA's trial terms say not to use it in production. source

Card: NVIDIA says Developer Program members get free access to NIM API endpoints for prototyping. It doesn't say whether a card is needed. source

Your prompts and code: Logged and may be used for training. NVIDIA's trial terms say it collects your prompts and outputs to improve its products, including AI models, and logs use for security and abuse monitoring. Don't upload personal or confidential data. source

“NVIDIA will collect the following data, without identifying specific users, to operate and improve the API Services and other products and services”
“User Content and Generated Content to improve NVIDIA products and services, including AI models.”

Company: NVIDIA Corporation, Santa Clara, California, United States source

Coding tools (per its own docs): NVIDIA doesn't publish setup guides for coding agents. Any tool that accepts a custom OpenAI-compatible base URL should work. OpenCode and Kilo Code also route some free models to NVIDIA's free endpoints, per their own docs. source

base_url = https://integrate.api.nvidia.com/v1 (OpenAI-compatible chat completions)

Official free page · Re-check: Model pages show a 'Free Endpoint' label and the trial rate limits. The list of free models changes; re-check weekly and before any post. 2026-10-30: added Mistral Large 4 (open weights released Oct 29, 2026; Mistral's own post links this NVIDIA page; page shows Free Endpoint).

• GitHub Models: Retired. GitHub's docs say it was fully retired on July 30, 2026. source
• Cerebras: Not a free tier: new accounts get $5 of credits that expire after 30 days, and a verified payment method is required. Cerebras says it has no permanent free tier. source
• Google AI Studio, Mistral and other model makers: These are the model makers' own free tiers, not third-party hosts. See the main free-tier list. free tiers

We read each host's free page, limits, terms and privacy policy. We list a model only if the host's own page says it is free. Posts on social media are leads only. Rows are re-checked at least every 7 days, and ended offers are removed.