Free LLM Models — Live-Tested, Updated Daily
Most “free AI models” lists are stale the day they are published: models get renamed, free tiers shrink, and a model that is listed as free may answer your agent with a 429 — or a 404. This page is maintained differently. An automated job re-reads the full OpenRouter catalog every day and keeps only models that are genuinely free and support tool calls — 16 out of 468 models qualify right now — and the top picks are probed with real API calls so you can see which ones actually answer.
Roster refreshed 2026-10-23 05:59 · Health check 2026-10-22 23:00 · Full data, scripts and agent recipes: github.com/XingMXTeam/freetoken-bots
Health check: what actually answered
One tiny real request per model (manual probes during initial setup: one chat completion per model, max_tokens=32, free-tier account). A 429 is a moment, not a verdict — free models share upstream pools.
| Model | Result | What happened |
|---|---|---|
| cohere/north-mini-code:free | ok | answered; then completed real Pi agent tasks (write + bash tools) with output verified by the maintainer |
| openrouter/free | ok | answered in ~1.3s, routed to poolside/laguna-s-2.1:free |
| google/gemma-4-31b-it:free | rate-limited | answered once, then HTTP 429 (upstream shared pool) on the next call minutes later |
| poolside/laguna-s-2.1:free | rate-limited | direct call HTTP 429, while the same model answered when reached through the openrouter/free router |
| nvidia/nemotron-3-super-120b-a12b:free | no-endpoint | listed in the catalog, but the chat call returned HTTP 404 |
The free roster: tool-calling models at $0
Every model below has all pricing fields at $0 and supports the tools parameter — the requirement for agent frameworks. Ranked with code-focused models first, then context size.
| # | Model ID | Context | Reasoning |
|---|---|---|---|
| 1 | poolside/laguna-s-2.1:free | 262K | yes |
| 2 | poolside/laguna-xs-2.1:free | 262K | yes |
| 3 | cohere/north-mini-code:free | 256K | yes |
| 4 | thinkingmachines/inkling-small:free | 1M | yes |
| 5 | thinkingmachines/inkling:free | 1M | yes |
| 6 | nvidia/nemotron-3-ultra-550b-a55b:free | 1M | yes |
| 7 | nvidia/nemotron-3.5-lightning:free | 1M | yes |
| 8 | nvidia/nemotron-3.5-super-vl:free | 1M | yes |
| 9 | dots-studio/dots-3-note-preview:free | 512K | yes |
| 10 | google/gemma-4-26b-a4b-it:free | 262K | yes |
| 11 | google/gemma-4-31b-it:free | 262K | yes |
| 12 | nvidia/nemotron-3-super-120b-a12b:free | 262K | yes |
| 13 | recursive/flint-v1.0:free | 262K | yes |
| 14 | xingchen-agi/xing4.0-29b-a4b:free | 262K | yes |
| 15 | nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free | 256K | yes |
| 16 | liquid/lfm-2.5-2.6b:free | 66K | yes |
Plus openrouter/free — a meta-router that picks a working free model server-side. Use it as the last-resort fallback in every failover chain.
Beyond OpenRouter: every free source, tracked
OpenRouter’s 50 requests/day runs out fast, so the same project tracks the other free sources. “Live” means the roster is pulled from the provider’s API; “seed” means a curated starter list for a provider whose catalog needs a key to read — labeled, never dressed up as live data.
| Provider | Kind | Free allowance | Roster |
|---|---|---|---|
| OpenRouter | Free tier | 50 free-model requests/day per account (free tier); models priced $0 | Live · 16 models |
| Google Gemini (AI Studio) | Free tier | Free tier, no card required. Community-measured: Flash models roughly 10 RPM and up to ~1,500 requests/day depending on model. Free-tier prompts may be used by Google to improve products; a project with billing enabled loses the free tier. | Seed · 3 models |
| Groq | Free tier | Free tier: headline ~1,000 requests/day per model, but a ~200K tokens/day cap (community tables) makes it run dry after a few long agent turns. Best for short tasks. | Seed · 5 models |
| NVIDIA NIM (build.nvidia.com) | Free tier | Developer program free credits: community tables list ~40 RPM and up to 10,000 requests/day per model - unusually generous; verify on your own account. | Seed · 3 models |
| Cloudflare Workers AI | Free tier | 10,000 Neurons/day free - roughly 4M input tokens/day on small models; the largest renewable free token pool, but you bring your own gateway/adapter. | Seed · 4 models |
| Zhipu GLM Flash (BigModel) | Free tier | Flash-series endpoints are free (rate-limited to ~1 concurrent request). Requires a real-name verified account (China). | Seed · 2 models |
| ModelScope (魔搭) | Free tier | Free inference API, community-reported ~2,000 calls/day. Requires Alibaba Cloud binding + real-name verification (China). | Seed · 3 models |
| SiliconFlow (硅基流动) | Free tier | A subset of models (mostly smaller Qwen/GLM/DeepSeek distillations) is free. Requires real-name verification (China). | Seed · 3 models |
| Mistral (La Plateforme) | Trial credit | Not a free tier: the Experiment plan includes monthly free credit (~$10 reported). Codestral is a strong free-ish coding option while credit lasts. | Seed · 2 models |
| Cerebras | Trial credit | Not a free tier: $5 trial credit, expires in 30 days, card required. Extremely fast inference while it lasts. | Seed · 2 models |
| Ollama (local) | Local | Truly unlimited: runs on your own machine - no key, no quota, no data leaving the device. Quality/speed bounded by your hardware. | Seed · 3 models |
| OpenCode Zen (DeepSeek free) | App-locked | deepseek-v4.1-flash-free is free for a limited time but only inside the OpenCode app - direct API calls are rejected (FreeTierError). Anonymous quota is shared per IP and was exhausted in the maintainer's test. | Seed · 1 models |
Use them in 3 steps
- Get a free key at openrouter.ai (Keys → Create Key). No credit card; then set a small spending limit on the key as a safety lock.
- Point any OpenAI-compatible client at
https://openrouter.ai/api/v1and use a model ID from the roster — only IDs ending in:free. The same model without that suffix is the paid twin. - Build failover: on failure, try the next roster model, then
openrouter/free. A single hard-coded free model will break — the health check above shows 429s and 404s happening to listed models within the same week.
Running an agent? The GitHub repo has a verified recipe for the Pi coding agent (config files included, write + bash tool tasks completed on a free model), with more agent recipes planned.
Frequently asked questions
Are these LLM models really free?
Yes. Every model in the roster has all pricing fields set to $0 on OpenRouter — prompt, completion, and every other meter. You need a free OpenRouter account and API key, but no credit card and no credit balance. Free models draw from a separate daily allowance (50 requests/day on a fresh account when this page launched).
Why does a 'free' model return 429 or 404?
Free models run in shared upstream pools. A 429 means the pool is busy right now — the same model often answers minutes later, or immediately when reached through the openrouter/free router. A 404 means the provider took the endpoint down even though the catalog still lists it. Both are normal on free tiers, which is why the health check above exists and why agents on free models need failover.
What is openrouter/free?
It is a meta-router, not a single model: OpenRouter picks a currently working free model server-side for each request. It is slightly less predictable than naming a model yourself, but far more reliable — the recommended last-resort fallback in any failover chain.
How do I keep 'free' from accidentally becoming paid?
Whitelist model IDs that end in :free. The same model without the :free suffix (for example poolside/laguna-s-2.1) is usually the paid twin, and most clients will happily bill you for it. As a second lock, set a small spending limit on your API key in the OpenRouter dashboard.
Can I use free models in my own product?
With your own key and within the provider's terms, yes — that is what the free tier is for. What you cannot do is pool keys or resell access to serve the public: free tiers prohibit it, and the quotas would not survive it anyway. Expect rate limits, and design for failover from the start.
How fresh is this list?
The roster is regenerated from the OpenRouter catalog every day by an automated job, and the health check probes the top models with real API calls. Every table on this page shows its own timestamp — trust the timestamp, not the page.
Data source: the FreeToken Bots open-source project (MIT), roster logic adapted from the FreeToken template by @MinLiBuilds. Free tiers and quotas change without notice — the timestamps on this page tell you how fresh each claim is.