Chronos-2, CLIP, and MobileNetV3: three production AI models for non-text workloads

A developer's analysis of the aiappdex.com model catalog highlights three non-text AI models seeing heavy production use: Amazon's Chronos-2 for zero-shot time-series forecasting (25.7M downloads), OpenAI's CLIP-vit-base-patch32 for zero-shot image classification (19.9M downloads), and timm's MobileNetV3-small-100 for lightweight image classification (17.4M downloads). The writeup notes Chronos-2 frames forecasting as language modeling with a T5 encoder-decoder, while CLIP's shared image-text embedding space enables hybrid retrieval, and MobileNetV3 targets hardware where larger models won't fit.

When I sort the model catalog at aiappdex.com https://aiappdex.com/ by total downloads, the 15th most downloaded model processes time-series data. The 14th handles zero-shot image classification. The 16th runs on hardware with under 3M parameters. None of them generate text. Three models from that tier deserve attention — not because they're new, but because the download numbers say someone is shipping them in production: | Model | Downloads | Task | |---|---|---| | Chronos-2 | 25,699,844 | time-series-forecasting | | CLIP-vit-base-patch32 | 19,936,700 | zero-shot-image-classification | | MobileNetV3-small | 17,428,712 | image-classification | Numbers from apps/ai-tools/src/data/models.json as of this writing. Chronos-2 https://aiappdex.com/models/amazon-chronos-2/ is Amazon's second-generation pretrained model for zero-shot time-series forecasting. The architecture frames the problem as language modeling: quantize the time-series values into tokens, run a T5 encoder-decoder, decode the output tokens back into predicted values. Apache 2.0. The zero-shot claim is the practically important part. Most forecasting approaches require labeled in-domain training data — you need historical examples for the domain you're forecasting. Chronos-2 was trained on a large corpus of diverse time-series and can generalize to new domains without fine-tuning. That's the same premise CLIP made for vision: collect enough diverse data at pretraining time and zero-shot transfer becomes viable. 25.7M downloads. In practice this means anomaly detection and demand forecasting in production data pipelines where someone had a time series and wanted a reasonable forecast without standing up a labeling operation. That use case — "I have the data, I don't have labeled examples" — is common enough to drive those numbers. What it doesn't do: outperform a well-tuned in-domain model on a domain where you have plenty of training data. Zero-shot is the alternative when you don't have that data, not a replacement when you do. CLIP-vit-base-patch32 https://aiappdex.com/models/openai-clip-vit-base-patch32/ is OpenAI's CLIP model with a ViT-B/32 image encoder, trained contrastively on 400 million image-text pairs. The training procedure aligns image and text representations in a shared embedding space — which means you can ask "which of these images is most similar to the phrase 'empty shelf'" without training a classifier for that specific concept. 19.9M downloads. 1,137 likes. The B/32 variant as opposed to ViT-L/14 trades accuracy for speed. In production that often makes sense: faster inference across more items is more useful than marginally better accuracy on fewer. The applications behind these numbers are image search, content moderation pre-filtering, and embedding images alongside text in retrieval-augmented generation pipelines. The cross-modal property is worth highlighting separately: CLIP's image encoder produces embeddings that are directly comparable to text embeddings from the same model. That's what makes it show up in hybrid retrieval systems — you can retrieve by text query and surface matching images without a translation step. MobileNetV3-small-100.lamb-in1k https://aiappdex.com/models/timm-mobilenetv3-small-100-lamb-in1k/ is a sub-3-million-parameter image classifier trained on ImageNet-1k via the timm library. At this size it runs on mobile hardware and embedded systems. 17.4M downloads. MobileNetV3 is in this list for a different reason than Chronos-2 or CLIP. It isn't zero-shot and it doesn't bridge modalities. It's a very small, very fast image classifier for hardware where nothing else fits. The downloads come from the overlap of two things: timm's consistent preprocessing API so it's easy to drop in as a swap for any other timm classifier and the large number of production systems that need a classification signal where a 10M-parameter model is already too large. The trade-off is explicit: you give up accuracy in exchange for a model that runs on edge hardware. For a use case where inference happens on device and latency matters more than peak accuracy, that's the right trade-off. The 103 likes low relative to downloads reflects that this is a utility model — people use it because it solves a specific constraint, not because they're enthusiastic about it. None of these are frontier models. None are mentioned in the AI news I read. All three are in the top 20 downloads on a catalog that includes hundreds of text-generation and instruction-tuned models. They're production workhorses in categories with genuinely different constraints than language generation. Chronos-2 targets the no-labeled-data case in time-series. CLIP targets cross-modal retrieval where text and images need to live in the same embedding space. MobileNetV3 targets inference on hardware where memory and latency are the primary constraints. If you're building something in one of those three spaces, 17–26 million monthly downloads is a better production-readiness signal than a benchmark table. It means the preprocessing API is stable, the common failure modes are documented, and someone has already debugged the deployment on your target hardware. Part of an ongoing 6-month experiment running three AI-curated directory sites. The technical claims here are real; this article was AI-assisted.