Tokener CA
Service online
Sign up
DocsChangelog

Changelog

A summary of version changes for the Tokener CA website, engine, and distribution programs.

Updated:

Tokener CA Website and Engine

  • 2026-06-01: Improved performance scoring and error classification; updated the Dify plugin marketplace link
  • 2026-05-29: Added machine names to the supply console and updated the performance benchmarking rules; Tokener CA is listed on the Dify marketplace, a news feed was added to the home page, and the Dify integration docs were updated to recommend using the Tokener CA plugin
  • 2026-05-22: Runtime model selection in the console, added iOS Web App support, improved iOS login redirect, session refresh, and cookie persistence, and added dedicated HuggingFace and OpenRouter endpoints
  • 2026-05-18: Fixed token speed calculation, improved the activity log display, added some hidden configuration buttons to the activity log, cached all query settings locally, and shipped several feature iterations for the Tokener CA Playground
  • 2026-05-14: In the supply console, performance scores can be clicked to trigger a manual retest, token consumption shows input and output speeds, added support for llama.cpp, and updated the developer integration docs and various third-party tool integration docs
  • 2026-05-13: In the Lab and Get console, added the Tokener CA Playground to test the various capabilities of platform models
  • 2026-05-11: provider v0.3.1 and above show the VRAM and context size of nodes
  • 2026-05-08: Added performance testing; nodes scoring above 60 get priority in task allocation, and nodes scoring 60-100 have equal weight in receiving tasks. When a node is offline for more than 1 minute, it automatically runs a performance test upon reconnecting to the platform. Improved task request error messages
  • 2026-05-06: Added node statuses: processing and pending; fixed several matching-engine scheduling hangs and frontend node-status hang issues
  • 2026-05-04: Improved node information; added the Changelog page
  • 2026-05-03: Optimized the matching engine's prefill intermediate frames to improve call stability; added the Relay lease feature, which greatly increases inference cache hits; added a minimum provider version requirement; added the Downloads page

Tokener CA Desktop

v0.5.3 (2026-08-17)

  • Added the qwen3.8:27b model

v0.5.2 (2026-06-01)

  • Limited the output size of image request logs to avoid log and memory bloat caused by large images or long payloads
  • Added compatibility for boolean-style reasoning parameter hints, improving call compatibility across different OpenAI-compatible clients
  • Limited the desktop Provider log buffer size to reduce memory growth during long-running sessions
  • Reduced the Provider log push frequency when the desktop window is hidden, lowering background overhead
  • Optimized log rendering and cleanup to prevent large volumes of logs from piling up and affecting UI responsiveness

v0.5.1 (2026-05-29)

  • Fixed runtime version detection for OpenAI-compatible runtimes such as LM Studio, prioritizing the local LM Studio application version to avoid mistakenly reporting the backend engine version as the application version
  • Improved runtime version detection and context window recognition compatibility, covering responses from different interfaces such as vLLM, SGLang, oMLX, and llama.cpp
  • Extended the timeout for OpenAI-compatible chat requests and raw passthrough requests to 10 minutes, reducing request interruptions caused by long inference or a slow first token
  • Added debouncing for the Relay disconnection offline state and state-generation validation, preventing stale offline updates from overwriting an already-reconnected state
  • Fixed the z-index of advanced configuration dropdown menus, preventing model, runtime, and other selection menus from being obscured by panels or advanced configuration items

v0.5.0 (2026-05-21)

  • Added desktop runtime model options, including options for third-party packaged and optimized images such as unsloth, allowing more precise model IDs to be selected across different local inference frameworks
  • Unified OpenAI-compatible capability preflight, performing stricter detection of text, image, tool, reasoning, and other capabilities at startup
  • Added runtime model ID preflight and retained the model context window information returned by the runtime
  • Added periodic capability rechecks, so capabilities can be re-detected after a node's model or framework capabilities change during operation
  • Fixed Relay disconnection reconnection and task-acceptance state recovery issues, reducing state desync after reconnecting
  • Recognized Provider Relay status logs, making node connection status messages more accurate
  • Reduced the image preflight prompt token budget to lower startup preflight overhead
  • Added a single-instance restriction on the desktop to prevent duplicate launches from causing multiple windows or Provider processes to interfere with each other

v0.4.1 (2026-05-18)

  • Added a UI scaling slider in the bottom-left corner
  • The Windows version now hides to the tray menu on close; right-click the tray icon to quit
  • Improved various UI messages and animations
  • Fixed 2 bugs where disconnection reconnection could cause permanent drop-offs; staying online is now more stable
  • Improved compatibility of image capability preflight at startup
  • Fixed inaccurate AMD GPU VRAM detection
  • Improved the security of OpenAI-compatible runtime error output

v0.4.0 (2026-05-15)

  • Changed the default inference framework from Ollama to LMStudio, and removed Others
  • Added llama.cpp framework support
  • Strengthened local model self-checks and max-tokens self-checks
  • Various server connection detail improvements

v0.3.1 (2026-05-11)

  • Added LMS tudio framework support
  • No longer specifically supports the KTansformers framework; it is now grouped under Others
  • Enforced detection of model context size
  • No longer supports a custom "heartbeat interval"
  • Added several pieces of model and hardware information reported by nodes
  • Improved Linux runtime stability and reduced crashes
  • Improved OpenAI-compatible protocol fault tolerance and stability

v0.3.0 (2026-05-07)

  • Supports multiple inference frameworks such as vLLM, SGLang, LM Studio, and oMLX
  • Added a portable Linux GUI build
  • Added API key format validation
  • Added a version update popup notification
  • The save feature now supports frameworks and color schemes
  • Changed the default concurrent inference to 1

v0.2.0 (2026-05-04)

  • Added a signed macOS .dmg build
  • Reports the desktop version to support node list monitoring.
  • Various frontend improvements and bug fixes.

v0.1.1 (2026-05-03)

  • Released the portable Windows .exe GUI build.

Tokener CA Supply: Command-Line Edition

v0.5.2 (2026-06-01)

  • Limited the output size of image request logs to avoid log and memory bloat caused by large images or long payloads
  • Added compatibility for boolean-style reasoning parameter hints, improving call compatibility across different OpenAI-compatible clients

v0.5.1 (2026-05-29)

  • Fixed runtime version detection for OpenAI-compatible runtimes such as LM Studio, prioritizing the local LM Studio application version to avoid mistakenly reporting the backend engine version as the application version
  • Improved runtime version detection and context window recognition compatibility, covering responses from different interfaces such as vLLM, SGLang, oMLX, and llama.cpp
  • Extended the timeout for OpenAI-compatible chat requests and raw passthrough requests to 10 minutes, reducing request interruptions caused by long inference or a slow first token
  • Added debouncing for the Relay disconnection offline state and state-generation validation, preventing stale offline updates from overwriting an already-reconnected state

v0.5.0 (2026-05-21)

  • Added options for third-party packaged and optimized images such as unsloth
  • Unified OpenAI-compatible capability preflight, performing stricter detection of text, image, tool, reasoning, and other capabilities at startup
  • Added runtime model ID preflight and retained the model context window information returned by the runtime
  • Added periodic capability rechecks, so capabilities can be re-detected after a node's model or framework capabilities change during operation
  • Fixed Relay disconnection reconnection and task-acceptance state recovery issues, reducing state desync after reconnecting
  • Reduced the image preflight prompt token budget to lower startup preflight overhead

v0.4.1 (2026-05-18)

  • Fixed 2 bugs where disconnection reconnection could cause permanent drop-offs; staying online is now more stable
  • Improved compatibility of image capability preflight at startup
  • Fixed inaccurate AMD GPU VRAM detection
  • Improved the security of OpenAI-compatible runtime error output

v0.4.0 (2026-05-15)

  • Changed the default inference framework from Ollama to LMStudio, and removed Others
  • Added llama.cpp framework support
  • Strengthened local model self-checks and max-tokens self-checks
  • Various server connection detail improvements

v0.3.1 (2026-05-11)

  • Added LMS tudio framework support
  • The ktransformers framework is no longer specifically supported; it is now grouped under Others
  • Enforced detection of model context size
  • No longer supports a custom "heartbeat interval"
  • Added several pieces of model and hardware information reported by nodes
  • Improved Linux runtime stability and reduced crashes
  • Improved OpenAI-compatible protocol fault tolerance and stability

v0.3.0 (2026-05-07)

  • Supports multiple inference frameworks such as vLLM, SGLang, LM Studio, and oMLX
  • Added a portable Linux .sh build
  • Changed the default concurrent inference to 1

v0.2.0 (2026-05-04)

  • Added a portable macOS .sh build
  • Reports the desktop version and handles control-plane version gating rejections, making it easier to prompt users to upgrade.
  • Optimized prefill and reasoning intermediate frames together with the engine to improve call stability.
  • Optimized the Relay lease feature together with the engine to greatly increase inference cache hits.
  • Fixed a UTF-8 output issue in Windows GPU detection.

v0.1.1 (2026-05-03)

  • Released the portable Windows .cmd build.