What’s New

The last 10 stable releases of WeightsLab, newest first. Each card links to its full release notes on GitHub, where the complete “What’s Changed” commit list lives.

Note

Only stable releases appear here. Every tag also publishes a vX.Y.Z.devN pre-release to TestPyPI for verification; those are build plumbing rather than user-facing changes.

Tip

Upgrade with pip install --upgrade weightslab. Setting up for the first time? Start at Quickstart.

October 20, 2026

v2.0.3

  • PR #312

    Summary

    This PR adds learned 2D/3D projections during training so model representations can be explored live in the Studio. By default, it uses a parametric UMAP trained alongside the model, allowing the projection to track representation changes as training progresses and to embed new samples without refitting. The projection can be configured per model, turned off when needed, and rendered alongside training progress in the Studio.

    Full description in the pull request.

https://github.com/GrayboxTech/weightslab/releases/tag/v2.0.3

October 1, 2026

v2.0.2

  • PR #306

    Summary

    • Improve experiment-view synchronization to operate incrementally instead of rebuilding the full dataset.

    • Reduce H5 update overhead and lock duration for large experiments.

    • Improve error reporting for failed synchronization and view-building operations.

    • Prevent training logs and tqdm output from appearing in unrelated notebook cells.

    • Fix communication issues caused by process-wide stdout/stderr redirection.

    • Fix the ‘weightslab se’ command for certificate generation.

    Full description in the pull request.

  • PR #287

    Summary

    • add a versioned, JSON-serializable model graph and per-layer/per-neuron inspection API

    • add validated dependency-aware helpers for add, prune, freeze, unfreeze, reset, and perturb operations

    • make freeze/unfreeze idempotent at the public API boundary

    • document the model graph ontology, dependency types, and editing workflow

    • add regression coverage for graph serialization, live structural edits, frozen state, and invalid targets

    Full description in the pull request.

https://github.com/GrayboxTech/weightslab/releases/tag/v2.0.2

August 26, 2026

v2.0.1

  • PR #300

    UI Fixes and upgrade docs

  • PR #299

    Add Opencode binary forgotten in v2

https://github.com/GrayboxTech/weightslab/releases/tag/v2.0.1

August 21, 2026

v2.0.0

What’s New ?

New Features

  • Runs management, unified UI to browse, organize, rename, and inspect experiment runs.

  • Error bands & outlier highlighting, curves now display statistical bands and visually emphasize anomalous steps.

  • Relabelling export, export tagged/annotated data to external tools (CVAT, V7, etc.) for downstream relabelling workflows.

  • Integrated OpenCode Agent, full agent loop support (code generation, training, monitoring, report creation) directly inside WeightsLab.

  • Multimodal data support, unified handling of images, videos, metadata, and structured signals.

  • Automatic resource monitoring, GPU/CPU/RAM usage tracked and surfaced during training and agent operations.

  • Dynamic HTML report generation, multi‑section experiment reports with plots, dataset analysis, training insights, and test results.

Fixes & Improvements

  • Agent stability improvements, better token management, reliable process detaching, consistent initialization, and workspace‑safe lifecycle.

  • Plotting upgrades

  • Correct zoom behavior across large step ranges

  • Bright color palettes in light mode

  • Outlier visualization improvements

  • Right‑click actions: BBS, highlight, hide curve, step notes, load weights, color changes

  • Signal pipeline fixes, improved decimation, preservation of special points, kernel stability, and classification logic.

  • DB performance improvements, safer handling of large histories, better compaction, and reduced memory pressure.

  • Tag painter fixes, more reliable tagging, discarding, and annotation workflows.

  • Workspace & session recovery, restart window reloads ongoing sessions, history, and conversation context.

  • UI polish

  • Search bar cleanup

  • Agent input bar sync

  • Regex‑based research plots

  • Updated sandbox modes

  • Improved multimodal previews

  • Cross‑platform testing, validated on Windows, Ubuntu, Jupyter, and Google Colab.

Developer Experience

  • Unified configuration, examples now rely on clean cfg files instead of hardcoded defaults.

  • Improved CLI, better agent commands, clearer /clear and /compact, stable loop behavior.

  • Changelog & documentation updates, new “What’s New”, migration notes (W&B / v51 / 3LC), updated examples, and expanded UI documentation.

Experimental & Advanced

  • Video generation workflows, multi‑input styles, real‑world models, and dataset‑driven video tasks.

  • Image generation workflows, PyTorch‑based generation paths integrated with agent prompts.

https://github.com/GrayboxTech/weightslab/releases/tag/v2.0.0

August 5, 2026

v1.5.0

  • PR #285

    • This PR introduces Google Collabs fixes and new integration:

    • Google Collab fixes

    • Jupyter Notebook Integration

    • Live notebook on-training allowing user to query, analyze, and ask agent to generate code to analyze ongoing experiment from the UI

    • Report generation with agent from code, CLI, or UI

https://github.com/GrayboxTech/weightslab/releases/tag/v1.5.0

August 4, 2026

v1.4.2

  • PR #283

    Summary

    • Wraps the “Install from TestPyPI/PyPI (wait until indexed)” + “Verify package import” steps in both test-install-from-pip-dev-release and test-install-from-pip-main in a single retried unit (nick-fields/retry@v3, max_attempts: 3, retry_on: any).

    • Guards against transient failures unrelated to the package itself: the registry hasn’t finished indexing a fresh upload yet, or a one-off network blip during install/import.

    Full description in the pull request.

https://github.com/GrayboxTech/weightslab/releases/tag/v1.4.2

July 24, 2026

v1.4.0

  • PR #266

    > Re-opened from #265 with the Ultralytics example notebooks excluded (kept on dev only, not ready for this release).

    WeightsLab v1.4.0, Release Notes

    🚀 Highlights

    • Collab compatibility improvements

    • PyTorch notebooks support

    • UL notebooks integrations

    • Dependency fixes (notably around numpy > 2 and protobuf)

    • Docker dependency cleanup

    • Removed Docker-related dependencies/repositories

    • Updated UI Docker bridge (including rename)

    • Updated README and docs accordingly

    Full description in the pull request.

https://github.com/GrayboxTech/weightslab/releases/tag/v1.4.0

July 10, 2026

v1.3.3

  • PR #250

    v1.3.3 Release

    Features & Improvements

    • Collab Integration: Bundled examples with Google Collab support, tunnel serving function, and documentation links

    • UI Enhancements: Fix grid size button, add restart action in dev mode, Firefox input arrows fix

    • Sorting: Default decreasing order (high loss values prioritized), configurable sort

    • CLI Upgrades: Match latest features—hyperparameter modification, experiment status check, data tagging/discard, agent queries

    Full description in the pull request.

https://github.com/GrayboxTech/weightslab/releases/tag/v1.3.3

July 7, 2026

v1.3.2

  • PR #243

    Fix histogram generation (previously interpreted numeric values as categorical values)

https://github.com/GrayboxTech/weightslab/releases/tag/v1.3.2

July 7, 2026

v1.3.1

  • PR #242

    • Improve agent:

    • Add new agent actions, e.g., dump model or architecture, data state, and modelling actions

    • Change default model used from 70B to “~google/gemini-flash-latest”

    • Optimize agent pipeline

    • Revamp the documentation (new widgets and content)

    • Allow user to generate categorical histogram from metadata

    • Improve CLI

    • New WeightsLab Banner

    • Improve, fix, and add new user weightslab commands (e.g., weightslab cli)

https://github.com/GrayboxTech/weightslab/releases/tag/v1.3.1

Every release on GitHub →