Pumas Library v0.6.0

This update resolves bugs downloading and displaying repos that have multiple models with shared names mixed together. It includes a significant performance improvement making Pumas Library effectively free to run in the background. There were also updates to make the APIs more consistent and inline with Transformers conventions and exposing core functionality.

Changes are underway in the inference server (Llamacpp, Ollama) to provide easy serving of LLMs with configuration for simultaneous models and device selection. It is not functional in this release, but it is scaffolded for the next release. Pumas Library will only provide basic inference options, for more advanced model support and production grade features use Pumas Library via Pantograph.

Added

• Selected artifact identity across download metadata, persisted package facts, frontend download state, migration dry-run reporting, and contract documentation
• Model-library update notification streams from the core update feed through RPC, Electron preload, and frontend subscriptions
• Backend download snapshot publishing and streamed model-download updates through RPC, Electron, and the React model manager
• Local runtime profile contracts, persistence, update streams, guarded profile operations, managed launch/stop commands, and profile-scoped Ollama model actions
• Managed Ollama and llama.cpp runtime profile support, including launch environment derivation, binary override preparation, provider routing, router catalog generation, presets, and dedicated llama.cpp profile launch flows
• Runtime profile settings and model route editors in the frontend, plus profile subscription hooks and route endpoint resolution
• Fast model selector snapshot contracts, index projections, direct API and read-only model-library access, batch selected-model hydration, and IPC/local-client selector APIs
• Local instance endpoint registration, ready-instance discovery, and local model-library update streaming for explicit local clients
• Status telemetry update streams bridged through Electron and consumed by the frontend

Changed

• Reworked API ownership so local clients are explicit and UniFFI API bindings are owner-only, removing stale utility, model, runtime, and migration client fallback branches
• Replaced several polling paths with pushed subscriptions for model-library updates, download state, runtime profiles, and status telemetry
• Routed Ollama list, load, and model actions through runtime profiles while preserving default singleton behavior
• Scoped active download indicators, local download rings, recovery state, and full-repository progress identity by selected artifact
• Preserved artifact slug separators and exposed planned migration action kinds for clearer migration review

Performance

• Cached system resource snapshots and app-liveness reads to reduce repeated status work
• Bounded backend runtime worker threads used by RPC runtime paths
• Lightened telemetry stream update payloads and removed the idle telemetry sampler
• Added selector snapshot timing coverage and local selector snapshot latency measurement

Pumas Library v0.5.0

The library database has been standardised around the HuggingFace Transformers conventions for easier compatibility with inference back-ends such as vLLM, MLX, PyTorch, Llamacpp, and others.

As of this release, Pumas Library has been adopted as the canonical source for models in Pantograph, the local AI framework for node-based concurrent multi model workflows that wont cause OOM errors. Pumas has various language bindings to integrate heedlessly into other apps, and has been validated in some small apps in past releases. But Pantograph is the first full scale deployment providing Pumas with real workload R&D as we build towards a V1.0.0 release.

In future releases we are looking to add a cryptographic secure layer for cloud API credentials to enable hybrid local + cloud uses. As well as changing the plugins which haven't seen much effort since January.

Added

• Package-facts APIs for summary snapshots, lazy package-facts resolution, model-reference resolution, update events, and persisted update-feed cursors
• Rich package evidence for import and execution review, including tokenizer diagnostics, special tokens, generation defaults, custom code requirements, auto-map and processor metadata, adapter/quantization/shard facts, sibling files, class references, source-repository evidence, and missing declared shards
• Hugging Face search and model-library compatibility hints for MLX, vLLM, unsupported backends, and unresolved/canonicalized Pumas model references
• Electron/RPC contract validation, plugin endpoint bridging, request schemas, empty-parameter handling, and preload drift tests
• Standards adoption, release-artifact, native-bindings, desktop-RPC, workspace ownership, and audit documentation, plus release-version, dependency-ownership, commit-message, README coverage, and file-size checks

Changed

• Refactored Rust runtime paths to move blocking filesystem and metadata work off async request paths, isolate path validation at API/RPC/UniFFI ingress, own background task handles, and cap RPC in-flight concurrency
• Split the frontend app shell, import workflow, model rows, version controls, mapping previews, metadata modals, install dialogs, download state, and API type contracts into smaller tested modules
• Split UniFFI bindings into focused API and FFI type modules while tightening native binding input validation
• Updated launcher and release flows to run from current build outputs without requiring toolchain dependencies in packaged release launches
• Updated README and workspace documentation for the current build, launcher, Rust, frontend, Electron, scripts, torch-server, and native-binding workflows

Fixed

• Hardened desktop IPC, RPC CORS, Torch sidecar access, LAN listener policy, path canonicalization, writable target probes, native import paths, migration/recovery targets, and direct file/open operations
• Restored frontend reliability for library download progress rings, inactive download indicators, cleaned metadata fallbacks, link-health rendering, app icon assets, model preview limits, failed shortcut rollback, and native accessible controls
• Corrected package-facts cache invalidation, metadata projection cleanup, library size calculation, proxy link exclusions, and unresolved library-path handling
• Fixed launcher dev-run backend artifact selection, release checksum digest exclusions, ComfyUI temp directory handling, Windows symlink test permissions, Rust/frontend CI failures, and write-target path canonicalization coverage

Pumas Library v0.4.0

Pumas Library 0.4.0 is a reliability and platform-hardening release focused on
making the desktop app, launcher workflows, bindings, and core library behavior much more dependable across Linux, Windows, and macOS.

Compared with 0.3.0, this release brings a stronger cross-platform desktop
workflow, a more resilient model-library core, better packaging and update
behavior, and significantly improved CI and contributor ergonomics.

Highlights

• Added a shared cross-platform launcher core for Bash and PowerShell with one
  consistent lifecycle contract
• Added bounded --release-smoke verification for release-facing desktop startup checks
• Added GitHub release detection and download prompts in the desktop updater flow
• Added packaged-app recovery for missing library detection, including persistent
• library-root selection
• Added C# UniFFI smoke coverage and packaging scripts for native/bindings
  release artifacts

Core Library Improvements

• Improved primary/client dispatch behavior so multiple processes can share one library more safely
• Hardened reconciliation and startup behavior to reduce unnecessary churn, stale
  state, and repair loops
• Improved canonical path handling and library-root normalization across
  platforms
• Improved model classification, task projection, and binding authority handling
• Added persisted metadata evidence to make download placement and local
  reclassification more reliable
• Added migration dry-run and audit tooling for library maintenance and metadata
  inspection

Desktop and Frontend Improvements

• Refined the desktop shell with better default window sizing and header behavior
• Improved packaged desktop startup and launcher-root discovery behavior
• Expanded model metadata editing and notes support
• Improved import, migration-report, mapping-preview, download, and version-
  management flows
• Refactored large frontend surfaces into smaller components and hooks with
  substantially better test coverage

Tooling, Packaging, and CI

• Moved workspace tooling to a pinned pnpm/Corepack ownership model
• Modernized CI bootstrap and workflow behavior for current GitHub Actions/
  runtime expectations
• Added stronger launcher verification paths in CI, including release-smoke
  coverage
• Improved release packaging for desktop artifacts and generated native/bindings
  artifacts
• Updated contributor, launcher, release, and repo documentation to reflect the
  current architecture and workflows

Stability Fixes

• Fixed UniFFI download request drift that was breaking pumas-uniffi builds
• Fixed Linux CI Electron startup issues under sandboxed runners
• Fixed macOS and Windows path-identity issues in Rust test expectations
• Fixed registry/test isolation issues that caused cross-test interference
• Fixed packaged-launch and shortcut issues that could break release usability
• Fixed Tokio example runtime configuration so documented examples compile with
  the pinned workspace features

Change Scope

From v0.3.0 to v0.4.0, the codebase changed substantially:

• 270 files changed
• 42,127 insertions
• 25,827 deletions

This release is less about one single headline feature and more about making
Pumas Library feel like a more coherent, release-ready platform: better startup
behavior, better desktop packaging, better update paths, stronger bindings
validation, and more trustworthy cross-platform workflows.

v0.3.0

Pumas Library v0.3.0

v0.3.0 is a focused upgrade over v0.2.0 that makes the library more capable for real-world import, download, and multi-process workflows.

This release adds first-class support for diffusers bundles across both Hugging Face downloads and local directory imports, with bundle-aware metadata, execution descriptors, and UI previews so imported models are easier to inspect before and after they land in the library. It also expands the frontend with richer model metadata, bundle component visibility, metadata columns, and better handling for classified directory imports.

Under the hood, v0.3.0 significantly improves library reliability. Partial downloads are now staged, resumable, and included in migration and integrity reporting. Model classification and metadata projection rely more consistently on SQLite-backed library state, Hugging Face hydration and bundle lookup are more robust, and recommended backend hints are now surfaced directly to clients. This version also introduces canonical reranker model type support and strengthens runtime dependency autobinding.

On the platform side, the release improves shared-library IPC by exposing library operations across processes and reusing the primary instance through UniFFI, while also fixing release validation, restricted-socket IPC test behavior, and cross-platform execution path issues for packaged builds, including Windows path handling.

Full Changelog: v0.2.0...v0.3.0

v0.2.0

Pumas Library 0.2.0 Release Notes

0.2.0 is a major platform release that moves Pumas from the early metadata flow in v0.1.0 to a full model-library + RPC + desktop stack with hardened contracts, better dependency semantics, and improved CI/release reliability.

Highlights

• Added a full SQLite-backed model library with indexing, search, migration tooling, and lifecycle APIs.
• Added metadata v2 foundations: classification, validation contract, review/reset flows, effective metadata APIs.
• Added Hugging Face enhancements: auth/token support, recursive tree fetch, grouped/shard-aware file selection, auxiliary config downloads.
• Added inference settings end-to-end (core types/defaults, API/RPC handlers, UI tab, UniFFI bindings).
• Added dependency profile pinning/compliance capabilities and then finalized a resolve-only dependency contract.
• Added migration report lifecycle: dry-run, execution, listing/pruning/deletion, and UI actions for opening JSON/Markdown reports.
• Added mapping/link exclusions and richer sync preview responses across core/RPC/Electron UI.

Breaking Changes

• Removed legacy metadata override endpoints (refactor(api)!).
• Dependency APIs were hard-cut to resolver-only semantics:
  • Pumas now resolves declarative requirements.
  • Consumer is responsible for environment checks/installs/runtime readiness.
• Metadata migration and dependency docs were updated to reflect the new contracts.

Reliability, Security, and Quality

• Improved download/import robustness:
  • interrupted download detection/recovery
  • early metadata indexing and HF quant sizing fixes
  • canonicalized model IDs
• Improved process/runtime stability:
  • removed async blocking paths for launch/stop
  • replaced unwrap-based log stream handling with IO-safe errors
• Electron safety hardening:
  • validated external URL scheme handling
  • ensured backend working directory exists before spawn
• Added regression/integration coverage in pumas-rpc.
• Cleaned and hardened CI workflow across Linux/macOS/Windows (actionlint pinning, caching fixes, optional native package handling).

UX / Frontend Improvements

• Better download telemetry (network status, active downloads, aggregate speed).
• Improved metadata object rendering.
• Added file-group multi-select for sharded downloads.
• Added mapping sync controls and link exclusion toggles.
• Added Hugging Face auth dialog with auto-401 detection.
• Added header minimize control and accessibility fixes.

Full Changelog: v0.1.0...v0.2.0

v0.1.0

Pumas Library v0.1.0

The first release of Pumas Library — a unified AI model management platform. Available as a desktop application for end-users and as a Rust library with bindings for six languages.

Pumas Library centralizes AI model storage, maintains consistency across applications, and eliminates duplicated models and scattered file management. Models are downloaded with rich metadata, full-text searchable, and automatically kept in sync.

Some QOL features are not yet available in the GUI. Headless packages are ready for use in integrated applications. Desktop app binaries are untested and provided as-is. If you identify a problem with them please open an issue on GitHub.

Highlights

• Unified model library — one location for all your AI models, shared across applications
• Full-text search — find models instantly across names, tags, and metadata (SQLite FTS5)
• HuggingFace integration — search, download, and track models directly from HuggingFace Hub with pause/resume support
• Application management — install, launch, and monitor ComfyUI, Ollama, PyTorch Server (Experimental)
• Real-time resource monitoring — CPU, GPU, RAM, and disk usage per-app and system-wide
• Model format conversion — convert between GGUF and Safetensors with multiple quantization backends
• Cross-process instance convergence — multiple processes share a single primary instance via local TCP IPC (JSON-RPC 2.0)

App and SDK

File	Description
pumas-library_*_amd64.deb	Linux installer (Debian/Ubuntu, x64)
Pumas-Library-Setup-*.exe	Windows installer (NSIS, x64)
Pumas-Library-*.exe	Windows portable executable (x64)
Pumas-Library-*.dmg	macOS disk image (Apple Silicon)
pumas-bindings-python-*.zip	Python bindings + native libraries
pumas-bindings-kotlin-*.zip	Kotlin/JVM bindings + native libraries
pumas-bindings-swift-*.zip	Swift bindings + native libraries
pumas-bindings-ruby-*.zip	Ruby bindings + native libraries
pumas-bindings-csharp-*.zip	C# bindings + native libraries
pumas-nif-elixir-*.zip	Elixir/Erlang NIF precompiled libraries
pumas-library-*.crate	pumas-library crate (for cargo install or vendoring)
checksums-sha256.txt	SHA-256 checksums for all release files