BERT-base-uncased, Electra, and why fill-mask models lead production NLP downloads

When I sort the model catalog in aiappdex by total downloads, the top of the list looks nothing like the AI discourse I read daily:

Numbers from apps/ai-tools/src/data/models.json as of this writing.

None of the top six are large language models. None of them generate text in a conversational sense. Number four is a model from 2018.

What fill-mask models are actually used for

fill-mask is the pipeline tag for masked language models like BERT, which are pretrained to predict a hidden token in context. In 2018 this was a breakthrough. In 2026 nobody blogs about it, but BERT-base-uncased still pulls 58.6 million downloads a month.

The actual usage is almost always fine-tuning. You take the pretrained checkpoint, attach a classification head, and train on your labeled dataset — customer intent classification, named entity recognition, content moderation labels. At 110M parameters, BERT-base runs on CPU without GPU infrastructure. Fine-tuning takes hours on a laptop. That's why it's in thousands of production ETL pipelines that nobody writes about.

The 2,910 likes suggest it has genuine fans, not just passive users who followed a tutorial once.

Electra: the discriminator that outperforms BERT-base on classification

Electra-base-discriminator sits at 56.9 million downloads — within rounding distance of BERT-base. Electra uses a different pretraining objective: instead of predicting masked tokens, it trains a discriminator to identify tokens that were replaced by a generator model. This produces representations that are more sample-efficient to train and, at the same parameter count, often score higher on classification benchmarks.

In practice the difference is subtle unless you're on a small dataset where the better pretraining generalizes further. Teams that have tried both tend to pick whichever initialized faster for their task and stick with it. The near-identical download counts suggest the user bases are largely overlapping.

Where text-generation fits in these numbers

Qwen3-0.6B sits at 21.4 million downloads — the highest text-generation entry I see in the top tier. It's a capable small language model, and 21M downloads is not a small number. But it sits below BERT-base by a factor of 2.7x.

Part of that gap is age: BERT has years of integrations, downstream checkpoints, and tutorials that compound the download count. But part is real economics. If the task is classifying 100,000 rows of support tickets into 12 intent categories, fine-tuned BERT on a CPU node is still cheaper and faster to serve than a generative model. The 21M downloads for Qwen3-0.6B reflect genuine adoption of small text-generation; the 58M for BERT reflect a different use case that generative models don't displace.

The embedding models above everything else

allMiniLM-L6-v2 at 246M downloads has no close competition. It's a sentence-similarity model from the sentence-transformers library that hit a minimum-viable size-quality tradeoff before anything else did — small enough to run on CPU, good enough to beat BM25 on most semantic search benchmarks.

BGE-small-en-v1.5 (65M downloads) and BGE-M3 (36.7M downloads) are the main challengers on quality, with M3 adding multilingual support and longer context. Switching costs and existing integrations keep allMiniLM-L6-v2 dominant even as newer models exceed it on MTEB benchmarks.

These embedding models are different from fill-mask models in one key way: you use them directly, not as a fine-tuning base. Give them a sentence, get a vector. That makes the integration simpler and the download curve steeper.

What this tells you about production NLP

The top download chart is a lagging indicator of what's deployed at scale, not what's experimentally interesting. Fill-mask and sentence-similarity models show up in this chart because teams found them worth integrating at production volume — not because they're generating benchmark coverage on leaderboards.

If you're building an NLP system and haven't compared fine-tuned BERT against API-based inference for your classification task, the 58 million downloads are a prompt to check. That count didn't accumulate by accident.

Part of an ongoing 6-month experiment running three AI-curated directory sites. The technical claims here are real; this article was AI-assisted.