We added a Specification section to every model page on ParaPulse

When you're evaluating a model on HuggingFace, you usually have to dig into the model card, check the repo files, or run the model locally just to find out basic architecture details — how many layers, what's the hidden size, what precision it runs at.

We just shipped a fix for that on ParaPulse.

What's new

Model detail pages now include a Specification section that surfaces the technical parameters directly — no repo digging required. If a model has configuration data available, you'll see things like:

- Architecture type and class

- Hidden size, layer count, intermediate (FFN) size

- Attention heads (Q and KV separately, so you can spot GQA/MQA at a glance)

- Max sequence length, vocabulary size

- Torch dtype (bfloat16 / float16 / float32)

- Activation function, RoPE theta, normalization epsilon

- MoE parameters (total experts, active experts per token) for models like Qwen3-MoE

It handles the full range of HuggingFace model types — transformers LLMs, BERT-style encoders, Whisper/audio models, timm vision models, CLIP, and ResNet-style CNNs — not just the big LLMs.

Why it matters

When you're choosing between five similar models, knowing that one uses GQA with 8 KV heads vs full MHA, or that one caps out at 32K tokens vs 128K, changes the decision. That information was always in the repo — it just wasn't surfaced anywhere useful.

The model list also now shows a small indicator icon next to models that have spec data available, so you know before you click.

Would love feedback on what other parameters would be useful to surface. There's a lot more in a typical config file than we show by default.