🧊Two days of free models that couldn’t finish anything
Moonstone, the app I run all my AI chats from, has a row of free cloud models next to the paid ones. MiniMax and GLM through NVIDIA, a free pick from OpenRouter, Gemini on a free key. The idea is simple. When the paid account runs out, or the job is long and boring, a free model picks it up.
For two days none of them could finish a thing. I opened fifteen free chats between October 26 and 27. Saying hi worked every time. The moment I gave one real work, like going through an email from my mom, cleaning up who I follow on X, or reviewing my Body Movin’ game, it died with a message saying the model froze or stayed busy too long.
⏱️It was my own timer
I had every request logged, so I went back through them. The models weren’t freezing. A few nights earlier I’d asked for a watchdog that stops a stuck model instead of letting me wait forever, and it stopped any free chat after 90 seconds of silence or 45 seconds of retrying.
Real work doesn’t fit in 90 seconds. GLM and MiniMax were taking 80 to 150 seconds a step and still getting answers back. NVIDIA’s free key also has a per minute limit, so when the agent fired a few tool calls quickly, NVIDIA said slow down, the bridge waited, and my timer counted the wait as a freeze. If the same limit hit three times, the bridge benched the whole provider for five minutes, which killed the chat on the spot.
The last one was the sneakiest. On the retest MiniMax handed part of the job to a helper agent, which can work quietly for up to fifteen minutes. Three minutes in, Moonstone saw no output and stopped it.
📏So I measured every one myself
I wanted real limits, not guesses, and the official model pages mostly wouldn’t load without a browser. The catalog sites that did load disagreed with each other. GLM’s answer limit was listed as 944 thousand tokens on one and 128 thousand on another. So I wrote a test that asks each provider directly. It checks how much the model can read by hiding a code in the middle of a huge pile of text and asking for it back, how long an answer it will accept, whether it uses tools correctly over several steps, which setting turns its thinking on, and how fast it writes.
Nemotron 3 Ultra on NVIDIA was the standout. It found the hidden code in a million token document in 50 seconds and wrote about 42 tokens a second. MiniMax M3.1 reads 256 thousand tokens on NVIDIA and refuses anything bigger, even though catalogs say a million, and it can answer up to 131 thousand. Nemotron VL on OpenRouter reads 128 thousand, writes about 162 tokens a second, and was the only model that got all four tool steps exactly right. DeepSeek V4.1 Flash also reads a million but takes over two minutes before its first word. NVIDIA retired GLM 5.3 the same morning I ran the test, and its Flash version timed out on every call. Gemini was answering “high demand” the whole time. NVIDIA’s free key gave me about 37 requests in the first 48 seconds before it started saying no.
🔧What changed
The bridge that sits between Moonstone and the free providers now keeps a live record of every request it has open, and Moonstone reads it. If a request is still out, or a helper agent or tool is still running on the computer, that counts as working and nothing gets stopped. A chat only gets stopped after eight minutes with nothing moving anywhere.
The bridge also paces NVIDIA calls to stay under that per minute limit instead of hitting it, so a busy chat waits a few seconds in line instead of failing, and Moonstone says it’s waiting on NVIDIA’s limit instead of calling it frozen. Each model got its real numbers too. Every free chat had been held to 4 thousand token answers and a 128 thousand token window no matter what the model could do. MiniMax can now hold far more of a job in memory, and every free model can write answers eight times longer. GLM and a Llama model that scored badly on my own work came off the menu.
🎮The real test
Then I gave MiniMax the exact job that had died on it the night before: go over my Body Movin’ game and make the body tracking better. It worked for about six hours across three sittings, 421 model steps. Once, early on, a single step grew so big it ran past a four minute cutoff, so I capped MiniMax at 160 thousand tokens of memory, which keeps every step to a couple of minutes. After that nothing cut it off. Slow is fine with me. Finishing is the part I care about.
It found something real. The phone’s tracker only sends about ten body frames a second while the TV draws sixty, so almost nine out of ten frames on screen showed the body frozen in place, and then it jumped. It also caught that my own test replay had been faking a smooth sixty frames a second, so every test I’d run on the game’s feel was measuring a feed that never happens. It fixed the replay, then made the drawn body glide toward where the tracker is heading instead of stepping. On my real session from that morning, frozen frames dropped from 88 percent to under 1 percent. It checked that the kick counting reads exactly the same data as before, frame for frame, and replayed my real boxing session to confirm the same ten kicks still land at the same moments.
A free model did that, on a free key, while I was doing other work. It just needed the room to finish.