I pull model stats weekly for aiappdex.com and five models keep surfacing for different reasons — either because their download counts put them in production-serious territory, or because their likes-to-downloads ratio tells a different story than the raw number suggests. Here's what I'm watching.
Kokoro-82M — TTS at 82 million parameters and 11.3 million downloads
Kokoro-82M shows 11,290,774 downloads alongside 6,806 likes. For a TTS model at 82 million parameters, that download count signals production use. Most TTS work defaults to cloud APIs; a model this small accumulating this much pull suggests developers are running it locally for latency or cost control.
I run a TTS pipeline for video narration in this project, currently using a hosted API. Kokoro is the model I'd evaluate first if I wanted to move that on-device. The parameter count is small enough to fit on edge hardware. I'm running a shelf detector on a Raspberry Pi 3 where the NCNN model is 36 MB; the Pi 3 clocks a median 8.5 seconds per image scan, which is fine at hourly cadence. TTS inference for a 60-second narration clip would be a different workload, and that's a case I haven't benchmarked — but the size class is plausible.
Amazon Chronos-2 — 25.7 million downloads, 426 likes
Chronos-2 is a time-series forecasting model from Amazon with 25,699,844 downloads and 426 likes. That ratio — very high downloads, minimal engagement — almost always means infrastructure use. Models people demo and share results from accumulate likes. Models embedded headless in data pipelines don't.
25.7 million downloads for a forecasting model is significant. General-purpose forecasting has been harder to commoditize than NLP or image generation — most real-world use is still domain-specific fine-tuning. A capable open-weight general forecaster reaching this download count suggests it's clearing the "good enough" bar for actual data pipelines, not just benchmarks.
gpt-oss-120b — OpenAI open-weight at 5.17 million downloads
gpt-oss-120b has reached 5,174,914 downloads with 5,148 likes. OpenAI releasing a 120B open-weight model was notable when it happened; tracking the download trajectory afterward is more useful than the announcement.
5.17 million downloads at frontier scale is not hobbyist traffic. It's running in inference clusters somewhere. I track this alongside DeepSeek-R1 (1,013,158 downloads in models.json) and Kimi-K3 (2,617,373 downloads) as a signal of how fast the open frontier is moving. The gap between closed and open frontier performance that looked durable eighteen months ago has compressed in ways the download numbers reflect before benchmarks do.
whisper-large-v3 — the ASR default at 4.9 million downloads
whisper-large-v3 sits at 4,908,520 downloads and 6,232 likes. It's the oldest model in this list by release date and still climbing.
What I find instructive about Whisper's trajectory is what "good enough, well-documented, and widely integrated" does to a market segment: it becomes the default. The ASR landscape has newer nominally-better options, but Whisper's ecosystem integration means switching costs have accumulated faster than the quality gap has grown. I use this as a frame for any model I consider embedding in the pipeline — if it gets load-bearing, what do the switching costs look like in two years?
Z-Image-Turbo — 5,197 likes on 665K downloads
Z-Image-Turbo from Tongyi (Alibaba's AI platform) shows the inverse pattern from Chronos-2: 665,197 downloads with 5,197 likes. High likes relative to downloads almost always signals hobbyist or creative adoption — people experimenting and sharing results rather than running it headless.
It's less covered than FLUX.1 or Stable Diffusion XL, which is part of why it's worth watching. Models that generate community interest before they generate documentation tend to be the ones worth tracking early. I'm not using it in the project yet, but the engagement-to-download ratio is one of the signals I pay attention to in the ETL because it often predicts download growth before the growth shows up.
All five models are browsable at aiappdex.com. Download and like numbers here are sourced from the project's HuggingFace ETL — real API numbers pulled weekly, not manually estimated.
Part of an ongoing 6-month experiment running three AI-curated directory sites. The technical claims here are real; this article was AI-assisted.
Top comments (0)