Importing a HuggingFace config.json

Load any open-weights model from the file that defines its shape.

Every model on the HuggingFace Hub ships a config.json beside its weights, and that file fully determines the architecture. NEURAX reads it directly, so working on a model that already exists does not mean rebuilding it block by block.

1. 1Press Import, right of the workspace tabs.
2. 2Press Open File and pick a config.json, or paste its contents into the box.
3. 3The badge above the box names the format it recognised.
4. 4Press Read File. A summary appears: depth, width, heads, vocabulary, and the structural choices it read.
5. 5Read the amber box, if there is one — it lists anything the file did not state.
6. 6Press Import. The architecture appears on the canvas and the analysis runs.

What it reads#

• Dense decoders — LLaMA, Mistral, Qwen, Gemma, Phi, OLMo, StableLM, Falcon, GPT-2, GPT-NeoX, OPT, Bloom.
• Encoders — BERT, RoBERTa, DistilBERT, ELECTRA and relatives, which get a classification head instead of an LM head.
• Mixture of experts — Mixtral, Qwen-MoE and DeepSeek-style configs, including shared experts.
• Multimodal — the language tower, from text_config. Vision and audio towers are not imported.

How accurate is an import?#

Page 19 of 33. This guide is in English for now. It is the same guide that is built into the NEURAX desktop app (the ? button, or F1). Found a mistake? Tell us.