anytimelime — the free AI field report

2026-09-10 · generated by a free model interviewing free models · anytimelime, formerly Project42 · get the engine →

What's free in AI right now — 2026-09-10

Morning’s big reveal: inclusionai/ling-3.0-flash-vl:free finally cleared the tool gate after months of silence, and SambaNova’s brand‑new deepseek-ai/deepseek-v4-pro-0813 is free but runs at an 18‑second crawl.

Running the free tier feels like picking a few reliable oddballs. The coder flavor cohere/north-mini-code:free still hacks a function in under a second (0.8s), dots-studio/dots-3-note-preview:free drops a quick note at 1.6s, and the financial flash inclusionai/ling-3.0-flash-fin:free (0.7s) and its health twin inclusionai/ling-3.0-flash-sante:free (0.8s) prove sub‑second response isn’t extinct. The visual flash inclusionai/ling-3.0-flash-vl:free (1.3s) is the new kid on the block, while liquid/lfm-2.5-2.6b:free (0.9s) still serves as a quick‑think backup. If you need raw compute, deepseek-ai/deepseek-v4-pro-0813 (18.4s) is free but you’ll want a coffee while it thinks.

For builders on free tiers the churn stays the same. The OpenRouter “:free” set is stable enough for a pipeline that can tolerate a second or two of latency, but the SambaNova wall means any hopes for Llama‑3.3, MiniMax, or the other gated giants are behind a payment method. The new flash model adds a vision hook without cost, while the 18‑second heavyweight reminds you that “free” doesn’t always equal “fast.” Keep scripts flexible and rotate before the next gate swings shut.

This week’s deep dig confirmed that the community isn’t yet publishing “AnytimeLime Endpoint” sections beyond the core site, so the measured table remains the only source of truth.

To run it yourself: clone github.com/fusionpoweredsoftware/anytimelime, set LIME_KEY to an OpenRouter key, run ./lime.sh – it pulls today’s verified roster and rotates across the passing models, never spilling your key anywhere but your own machine.

Today's passing roster

The measured table

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

latencymodelvendortool callfree nowagentic
0.201-ai/yi-largenvidiauntestedDOWN—
0.8cohere/north-mini-code:freeopenrouteryesFREE3/5
0.3DeepSeek-V3.1sambanovauntestedCARD—
0.1adept/fuyu-8bnvidiauntestedDOWN—
1.6dots-studio/dots-3-note-preview:freeopenrouteryesFREE3/5
0.3DeepSeek-V3.2sambanovauntestedCARD—
0.1ai21labs/jamba-1.5-large-instructnvidiauntestedDOWN—
0.2google/gemma-4-26b-a4b-it:freeopenrouteryes (past)DOWN—
0.3Meta-Llama-3.3-70B-InstructsambanovauntestedCARD—
0.1aisingapore/sea-lion-7b-instructnvidiauntestedDOWN—
0.2google/gemma-4-31b-it:freeopenrouteryes (past)DOWN—
0.3MiniMax-M2.7sambanovauntestedCARD—
0.1bigcode/starcoder2-15bnvidiauntestedDOWN—
0.7inclusionai/ling-3.0-flash-fin:freeopenrouteryesFREE3/5
0.3MiniMax-M3sambanovauntestedCARD—
0.2databricks/dbrx-instructnvidiauntestedDOWN—
0.8inclusionai/ling-3.0-flash-sante:freeopenrouteryesFREE3/5
0.3gemma-4-31B-itsambanovauntestedCARD—
0.1deepseek-ai/deepseek-coder-6.7b-instructnvidiauntestedDOWN—
1.3inclusionai/ling-3.0-flash-vl:freeopenrouteryesFREE3/5
0.3gpt-oss-120bsambanovauntestedCARD—
60.1deepseek-ai/deepseek-v4-flash-0731nvidiayes (past)DOWN—
0.9liquid/lfm-2.5-2.6b:freeopenrouteryesFREE3/5
18.4deepseek-ai/deepseek-v4-pro-0813nvidiayesFREE3/5

7 made a tool call, 0 answered in text only, of 24 discovered. A fail only means the endpoint wasn't free to use right then — it demanded a payment method, or it was down or erroring — so tool-call ability stayed untested, not disproven; a model that passed in a past report keeps its yes. Numbers are single-run measurements from one location — your mileage will vary. Prose by cohere/north-mini-code:free; the table is measured, never model-written. Reproduce with ./free-scan.sh from the engine repo.