Inspect

Deep analysis of model architecture, memory layout, capabilities, tensor composition, and runtime compatibility

Inspect (02)

Inspect provides deep analysis of a loaded model — 3D architecture visualization, memory breakdown, capability detection, quantization distribution, runtime compatibility, and more.

A model must be loaded via the Load module before Inspect is available.

Isometric 3D Visualization

An interactive isometric view renders the model as stacked blocks:

• Embedding layer at the bottom
• Transformer layers stacked vertically, colored by attention/MLP tensor ratio
• Output layer at the top

Control	Action
Hover	Tooltip with tensor breakdown
+ / - buttons	Zoom in / out
Mouse wheel	Zoom
Reset button	Reset view

Memory Distribution

Six-component breakdown showing how memory is allocated:

Component	Description
Embeddings	Token embedding weights
Attention	Q, K, V, O projection matrices
MLP	Gate, up, down projection matrices
Norms	RMSNorm / LayerNorm weights
Output	LM head / output projection
Other	Miscellaneous tensors

Each component shows exact byte count, percentage, and a proportional bar.

Quantization Breakdown

For each dtype present (F32, F16, BF16, Q8_0, Q4_K_M, etc.):

• Tensor count, total size, percentage, and visual bar chart

Capability Detection

Analyzes model architecture to detect 7 capabilities with confidence scores:

Capability	What It Detects
Tool Calling	API/function calling ability
Reasoning	Chain-of-thought reasoning
Code	Code generation/understanding
Mathematics	Mathematical reasoning
Multilingual	Multi-language support
Instruction	Instruction following
Safety	Safety/alignment layers

Runtime Compatibility Matrix

Checks support across 8 popular inference runtimes:

Runtime	Formats
llama.cpp	GGUF
Ollama	GGUF
LM Studio	GGUF
KoboldCpp	GGUF
GPT4All	GGUF
Jan	GGUF
LocalAI	GGUF
text-generation-webui	GGUF, SafeTensors

Status: COMPATIBLE, PARTIAL, or NOT SUPPORTED.

Attention Architecture

• Query heads and KV heads count
• Head dimension and GQA ratio
• Visual head diagram

Tokenizer Info

• Tokenizer type (BPE, Unigram, WordPiece)
• Vocabulary size
• Special tokens (BOS, EOS, PAD, UNK) with IDs

File Verification

Click COMPUTE HASH to calculate SHA-256 fingerprint for integrity verification.

Data Export

• JSON — Full model metadata as JSON
• CSV — Tensor list with names, dtypes, shapes, and sizes

Layer Hierarchy

Expandable list per layer showing attention, MLP, norm, and other tensors. Filter by name, dtype, or layer range.

Tensor Browser

Searchable, filterable list of all tensors with:

• Tensor name
• Data type
• Shape dimensions
• Memory size