anytimelime — the free AI field report

What's free in AI right now — 2026-09-01

Seven models made a real tool call today, up from three yesterday. Read that sentence twice, though, because five of the seven — cohere/north-mini-code:free, dots-studio/dots-3-note-preview:free, google/gemma-4-31b-it:free, inclusionai/ling-3.0-flash-fin:free, and liquid/lfm-2.5-2.6b:free — were the entries we burned our own rate limit on yesterday and never reached. Today's roster is partly a measurement artifact, not a capability surge. The genuine veterans who passed again are minimax/minimax-m2.7:free and deepseek-ai/deepseek-v4-flash-0731.

The fast cluster is genuinely good news: inclusionai/ling-3.0-flash-fin:free at 0.9s, liquid/lfm-2.5-2.6b:free at 1.0s, and google/gemma-4-31b-it:free at 1.2s all cleared tool calling under two seconds. If you need a code brain, cohere/north-mini-code:free is the slowest passer at 4.3s but it's the only one wearing a "coder" label and it actually called the tool. The patience-tester is still deepseek-ai/deepseek-v4-flash-0731 — 12.9s today, an improvement from 58.5s yesterday, though it switched to an nvidia host so the comparison is apples to oranges. One warning worth flagging: minimax/minimax-m3:free slid from PASS yesterday to WEAK today — it answered in fluent text and silently ignored the tool schema. And SambaNova's payment wall is unchanged: seven models, one identical message demanding a payment method on file.

How to use them with AnytimeLime

Clone the engine from github.com/fusionpoweredsoftware/anytimelime, set LIME_KEY to your OpenRouter key, and run ./lime.sh. It fetches today's roster and rotates across the passing models so no single free tier absorbs all your traffic. Your key stays on your own machine — the website never asks for it.

If you already have an OpenRouter key, the model IDs above work directly with any OpenAI-compatible client. No AnytimeLime needed if you prefer to wire it up yourself.

The churn

Free rosters change weekly — models get gated behind payment walls, suffer silent provider errors, or disappear entirely. Today seven passed; last time this ran, three did — with six more benched untested because we'd burned our own rate limit before reaching them. Five of those six cleared the bar today, so yesterday's "three" was a floor, not a measurement.

Today's passing roster

• inclusionai/ling-3.0-flash-fin:free — openrouter · 0.9s · agentic 3/5
• liquid/lfm-2.5-2.6b:free — openrouter · 1.0s · agentic 3/5
• google/gemma-4-31b-it:free — openrouter · 1.2s · agentic 3/5
• dots-studio/dots-3-note-preview:free — openrouter · 1.8s · agentic 3/5
• minimax/minimax-m2.7:free — openrouter · 3.0s · agentic 3/5
• cohere/north-mini-code:free — openrouter · 4.3s · agentic 3/5
• deepseek-ai/deepseek-v4-flash-0731 — nvidia · 12.9s · agentic 3/5

The measured table

Click any model for its endpoint and ready-to-run snippets in curl, Python, Go, Rust, and more.

7 made a tool call, 1 answered in text only, of 24 discovered. Numbers are single-run measurements from one location — your mileage will vary. Prose by dots-studio/dots-3-note-preview:free; the table is measured, never model-written. Reproduce with ./free-scan.sh from the engine repo.