Skip to content

Tensormorph

An IDE for Model Weights

Instant to Open

Live Runtime Tracing

Catch Anomalies Fast

One Address Space

Any Compute Backend

3D, 2D & 1D Views

Model Morphing

Local or Hub

Tensormorph is a professional Tensor IDE for visually exploring, inspecting, comparing, debugging, and transforming tensors and open-weight AI models. It helps machine-learning engineers navigate models from complete architectures through layers, modules, tensors, matrix tiles, and individual scalar values. Tensormorph aims to make model weights easier to understand through interactive visualization, statistics, heatmaps, model comparison, quantization analysis, and scalable tensor inspection.

✅

Every view - 3D, 2D, 1D, or a single scalar - reads the same shared address space, so switching how you look at a tensor never means losing your place.

Open, navigate, inspect, compare, debug, and version model checkpoints - from a few million parameters to trillion-parameter Mixture-of-Experts models - without needing to fully load a model into RAM or VRAM to work with it. If a debugger, a profiler, and an IDE had a shared tool for weight space instead of source code, this is roughly it.

Get Started · Core Concepts · GitHub Repository (opens in a new tab)

Query weights like data

TQL (Tensor Query Language) is one grammar for selection filters, breakpoint conditions, and shareable URLs alike - plain ASCII that reads directly off precomputed tile statistics, so it stays fast even on a very large model:

# Every attention tensor with an unusually large value
layer.type == "attention" and tensor.absmax > 10

# Break the moment any tensor produces a NaN
tensor.nan_count > 0

Type it into the Console for a full REPL, prefix a filter with : from the command palette (Ctrl+P), or save a query once and reuse it across sessions. See TQL: Expressions and Queries.

One address space, every scale

Everything you inspect lives on the same ladder - model, architecture, layer, module, tensor, tile, row/column, down to a single scalar (see Model, Architecture, Layer and Module) - and every view is a different lens on that same address space, kept in sync:

  • 3D views for spatial structure: Architecture, Layer Stack, Tensor Volume, MoE Expert Galaxy, Parameter Map, Diff, Morph, Runtime Flow.
  • 2D views for precise per-element reading: the Matrix and Heatmap Editor.
  • 1D views for curves and distributions: the Signal and Distribution Editor.
  • The Scalar Inspector, for a single number in full detail.

Named workspace presets - Architecture, Compare, Morph, MoE, Runtime, Performance, and others - switch this whole layout at once for a given task, rather than requiring you to rearrange panels by hand.

What it isn't

🚧

Tensormorph doesn't bundle its own version control. Lob, a separate tensor-native VCS, is what actually stores weight history - Tensormorph anchors to it via commit pins rather than reimplementing it. Tensormorph also isn't primarily a training tool: it reads checkpoints and runtime traces rather than running training loops itself. And it hasn't shipped a public release yet - see Install the Desktop Application for what's confirmed so far.

Where to go next