VRAMLab

Local .gguf & config.json Binary Header Inspector (Zero Upload)

Drop any 4 GB to 200 GB .gguf weight file or Transformer config.json right from disk. VRAMLab slices only the leading 16 MiB in local browser memory via DataView to extract exact layer counts, GQA/MLA dimensions, context limits, and quantization types in under 5ms—zero bytes ever leave your machine.

Select Local Model File or Load Verified Sample Header

Reads only bytes 0..16,777,216 via File.slice() — multi-100GB .gguf files open instantaneously without copying full weights into RAM.

Or test immediately with synthetic binary headers:

Parsed Qwen3-32B-Q4_K_M.gguf (GGUF v3) in local browser memory — 0 bytes uploaded.

Extracted Architecture & Quantization Metadata

Qwen3-32B-Instruct · MOSTLY_Q4_K_M
Extracted Architecture & Quantization Metadata
Header Key Extracted Value
general.architecture qwen3
general.name Qwen3-32B-Instruct
gguf.version v3
gguf.tensor_count 707
gguf.quantization MOSTLY_Q4_K_M
qwen3.block_count 64
qwen3.embedding_length 5120
qwen3.attention.head_count 64
qwen3.attention.head_count_kv 8
qwen3.context_length 40960

Instant VRAM Projection from Extracted Header

RTX 4090 24GB
28.28 / 24 GB
Out of Memory (OOM) · OOM
RTX 5090 32GB
28.28 / 32 GB
Tight Fit (>90% Usable) · 39.4 tok/s
Mac Studio 128GB
27.66 / 128 GB
Fits Comfortably · 11.3 tok/s
NVIDIA H100 80GB
28.28 / 80 GB
Fits Comfortably · 77.9 tok/s
Open Full Breakdown in VRAM Calculator → Generate vLLM / Ollama Config → Models →

How Zero-Upload GGUF v2/v3 Binary Parsing Works

The GGUF binary container stores all model metadata immediately after the 4-byte ASCII magic ("GGUF" / 0x46554747), 32-bit version integer, 64-bit tensor count, and 64-bit KV pair count—well ahead of multi-gigabyte quantized tensor blocks. By slicing only the first 16 MiB with the browser File API, VRAMLab decodes UTF-8 keys, typed scalars, and tokenizer array lengths without reading tensor payloads.

For FP16, BF16, FP8, or AWQ checkpoints distributed as safetensors shards, dropping the accompanying config.json extracts hidden_size, num_hidden_layers, num_key_value_heads, kv_lora_rank, n_routed_experts, and sliding_window_pattern so you can size vLLM or SGLang clusters before downloading hundreds of gigabytes of weights.

VRAMLab evaluates each model from its verified transformer geometry—branching explicitly across Multi-Head Attention (MHA), Grouped-Query Attention (GQA), DeepSeek Multi-head Latent Attention (MLA), and hybrid local/global sliding-window stacks—so capacity plans match real vLLM, SGLang, Ollama, and llama.cpp allocations.