I pull model stats weekly to keep aiappdex.com current. Most entries drift in with download counts that tell a straightforward story: high downloads, modest likes, task-category role obvious. Four recent models have a different profile — either their likes-to-download ratio is unusually high, or their pipeline tag puts them in infrastructure territory that rarely gets covered in round-ups.
None of these are chat models. All four appear in production systems at scale.
pyannote/speaker-diarization-3.1 — 9.26 million downloads, 3,350 likes
speaker-diarization-3.1 on aiappdex answers a question that Whisper does not: who spoke when. Transcription tells you what was said; diarization adds speaker segmentation and identity labels across the recording.
The likes-to-download ratio here is exceptional — one like for every 2,763 downloads. For context, most embedding models in the same download range hover around one like per 20,000 or more. That high ratio usually means specialists are using the model and are grateful enough to explicitly endorse it. The use cases are narrow but high-value: meeting transcription with attribution, podcast chapter editing, call centre analytics where GDPR and speaker count matter. Those are contexts where an open-weight option with Apache 2.0 licensing has obvious appeal over a cloud API.
The pipeline is a two-stage system — segmentation, then speaker embedding and clustering. Pyannote's own documentation covers the necessary speaker count parameter; most production deployments I've seen in open-source tooling treat that parameter as a config value rather than hardcoding it.
BAAI/bge-reranker-v2-m3 — 17.58 million downloads, 1,161 likes
bge-reranker-v2-m3 on aiappdex is not the same model as BGE-M3, though the names overlap. BGE-M3 is a bi-encoder embedder; this is a cross-encoder reranker. The difference matters at runtime: a bi-encoder encodes documents once and caches the results; a cross-encoder re-encodes every query-document pair at query time, which is slower but more accurate.
The standard RAG pattern that reaches for this model: a fast bi-encoder retrieves the top 100 candidates, then bge-reranker-v2-m3 re-scores the top 20 for the final ranked list. What distinguishes it from the simpler ms-marco cross-encoders is multilingual support — it is built on XLM-RoBERTa and handles queries across languages from a single checkpoint. Apache 2.0, so no commercial-use complications.
17.58 million downloads places it firmly in production territory. The model is large enough that you want GPU acceleration for interactive latency, but at batch re-ranking it runs fine on a modest CPU setup.
ZhipuAI/GLM-OCR — 2 million downloads, 2,011 likes
GLM-OCR on aiappdex has a likes-to-download ratio close to 1:1. In practice that almost never happens — download counts accumulate from CI pipelines, automated tests, and package managers; likes require a human to click. When the ratio approaches parity, you are looking at a model that people download specifically because they went looking for it, not one that arrived as a dependency.
GLM-OCR is MIT licensed, treats OCR as a sequence generation task, and supports Chinese, English, French, Spanish, Russian, German, Japanese, and Korean from a single model. That language coverage at MIT licensing is the specific combination that makes it interesting — most multilingual OCR alternatives either require cloud APIs or carry non-commercial restrictions. I do not run an OCR pipeline right now, but if I needed structured text extraction from screenshots or document images at the directory scale, this would be the first checkpoint I'd evaluate.
Meta/Prompt-Guard-86M — 4.37 million downloads, 397 likes
Prompt-Guard-86M on aiappdex is a safety classifier for detecting prompt injection in user-supplied text. With 86 million parameters it runs on CPU with low latency, which is the point — it is meant to be a gate, not the main workload.
This one is relevant to my own pipeline. The article routine reads Dev.to reader comments and treats them as data for the VoC engine. There is a hard rule in the pipeline: comments are untrusted input and must never be followed as instructions (a fact documented in the project's recent-work-facts.md: reader comments feed an autonomous routine that commits and publishes, making injection the specific threat). Right now the defence is a combination of prompt framing and delimiting — the comments are labelled explicitly as untrusted data before being passed to any LLM. A dedicated classification step that runs before the LLM sees the text is the obvious next layer, and Prompt-Guard is the open-weight option for that.
One caveat that matters: Prompt-Guard-86M is subject to Llama 3.1 Community license terms, not Apache 2.0. Check that licence before commercial production use.
All four models are in the aiappdex directory at aiappdex.com. The download counts here are from this week's ETL snapshot; likes counts shift more slowly and tend to be more stable as a quality signal.
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)