Read a model’s tokenizer.
safe · secure · no server · works offline · fast
…or just start typing at the cursor.
safe · secure · no server · works offline · fast
…or just start typing at the cursor.
Drop a .gguf and OmniViewer reads its embedded
tokenizer: the tokenizer model, the vocabulary
size, the special token ids (BOS, EOS, UNK, PAD) and the
chat template the model expects for its prompts. It reads
these from the header only, so it’s instant and entirely local.
The chat template is the exact Jinja string a runner uses to format a conversation for the model — useful to see before you wire the model into a chat UI. The same drop also shows the architecture, metadata and tensors.
It’s one lens on OmniViewer’s GGUF Viewer — Inspect LLM Model Files: drop a file and every tab is one click away. OmniViewer opens every file format, entirely in your browser — no upload, no server.
No. OmniViewer is a static page with no server-side processing: your .gguf is read directly by your browser, and every tab — architecture, metadata, tensors, tokenizer and stats — runs locally in a Web Worker. The model never leaves your computer.
GGUF keeps all of its metadata and its tensor table at the front of the file, before the weights. OmniViewer reads only that header prefix — a few megabytes at most — and never reads the multi-gigabyte weight body, so a 40 GB model opens as fast as a small one.
A lot, with no description file: the architecture, the parameter count (computed from the tensor shapes), the quantization and effective bits-per-weight, the number of layers, the embedding and context length, the attention type (multi-head, grouped-query or multi-query), the Mixture-of-Experts configuration, the RoPE settings and the embedded tokenizer including its chat template.
Yes. When the header carries an expert count (as Mixtral and other MoE models do), the architecture card shows how many experts the model has and how many are routed per token — for example, 8 experts with 2 active.
Bits-per-weight is the average number of bits each parameter takes on disk — a direct measure of how aggressively the model is quantized. OmniViewer computes it as the file size in bits divided by the parameter count (summed from the tensor shapes), so a Q4_K model comes out around 4.5 bits per weight.
All the common GGML tensor types, including F32, F16 and BF16 and the quantized types Q4_0, Q4_K, Q5_K, Q6_K, Q8_0 and the IQ family. The Tensors tab shows the exact mix, and the Architecture card names the dominant one.
Yes. OmniViewer parses the GGUF structure in pure JavaScript, so it works in any modern browser on Windows, Linux, macOS or a phone. No llama.cpp, no Ollama, no Python, no install.