Skip to main content

🧚 New models, self-checking training, and your workspace, your rules

Released on October 26, 2026

Two powerful new models join the family — TF K3 Optimised and Qwen 3.8 2.4T A95B — alongside reinforcement learning tuning that teaches your agents to check their own work, workspace model governance so admins control which models their teams can use, and your kanban boards, channels and routines on mobile. 🚀

🤖 New Models​

TF K3 Optimised — frontier intelligence, minus the wait​

Our new efficiency-tuned flagship takes the frontier K3-class architecture — a 2.8T-parameter multimodal model with native vision, a 1M-token context and interleaved thinking — and optimises its reasoning behaviour to think substantially shorter before it answers. The result is roughly 40% fewer reasoning tokens at comparable quality, which means faster answers and lower costs on coding, research and agentic work, with vision and function calling fully supported.

Qwen 3.8 2.4T A95B — built for the long haul​

Alibaba's most capable Qwen model to date is a 2.4T-parameter sparse mixture-of-experts with around 95B active parameters, designed for autonomous, long-horizon work: multi-day coding runs, research-paper reproduction and self-improving agents. When a task needs stamina as much as brilliance, this is the model to reach for.

Key Features:

  • TF K3 Optimised - frontier quality with roughly 40% shorter thinking traces, native vision and a 1M-token context
  • Qwen 3.8 2.4T A95B - a long-horizon specialist for multi-day coding and research work
  • Pick your tier - both models are available across our service tiers, so you can balance speed, cost and capability per task

🎯 Reinforcement Learning Tuning​

Agents that grade their own homework​

Fine-tuning just learned a new trick. With RLVR tuning (reinforcement learning from verifiable rewards), you describe a task and a checker that validates the outcome — and your model learns by attempting the task, getting scored against the checker, and improving with every round. It is a fundamentally different way to train: less copying examples, more earning the right answer.

The whole flow lives where your other tuning work already does. A new RL task document type gives you a conversational editor with dedicated task and checker blocks, the import wizard accepts RL task files with built-in validation, and a new RLVR option in your tuning configuration lets you set generations, tool calls, entropy and any extra packages your sandbox needs.

Key Features:

  • RL task documents - describe tasks and checkers in a purpose-built editor, and import RL task files with validation
  • RLVR tuning type - configure generations, tool calls and entropy directly from your tuning setup
  • Reliable by design - each training run is isolated per rollout, with support for extra packages your RL sandbox needs

🔐 Workspace Model Governance​

Your models, your rules​

Administrators can now curate the model catalogue per workspace. Choose exactly which models your teams can and cannot use, and the list is enforced everywhere — the web app, TF Code and the API all see the same governed catalogue, so there are no surprises. Your own fine-tuned models are always exempt, and the picker is colour-coded by model size tier to make the right choice at a glance.

Key Features:

  • Blocked models list - remove models from a workspace's catalogue, enforced across the app, TF Code and the API
  • Fine-tunes stay visible - your own custom models are never blocked
  • Clearer picker - models are colour-coded by size tier, in every language we support

🖥️ Your Workspace on Mobile​

Kanban, channels and routines in your pocket​

The mobile app grows up this release. Your kanban boards, channels and routines are now fully at your fingertips on the go — create and move tasks, chat with your channels and manage your recurring work, all from your phone.

Connecting to a remote device is simpler too. Mobile peering is now public and automatic: your relay is derived from your region, so there is no address to type — just pair and go.

Key Features:

  • Kanban on the go - create boards and tasks, and move work forward from your phone
  • Channels and routines - keep conversations and recurring schedules at hand wherever you are
  • One-tap peering - pair with a remote device without typing an address

🔌 Media Tools for Your Automations​

Voices, images and video — now through MCP​

Your agents and automations can now create media directly. New MCP tools let you list, clone and delete voices, generate speech from text, create images and turn prompts into video — so an automation can narrate its own report, illustrate a finding or produce a clip without anyone touching a dashboard.

📖 API Documentation: apidocs.toothfairyai.com - Media

🎯 TF Code: A Built-in Reverse Engineer​

Understand any shipped software​

A new reverse engineer command joins your local agent toolkit. Point it at an app bundle, binary, APK, WebAssembly module or a live URL, and it explains how the feature works — with evidence. It runs as a local agent in TF Code, so it is available wherever you work: the UI, a one-shot tfcode run command, or headless in your scheduled automations.

Key Features:

  • Any target - analyse apps, binaries, APKs, WebAssembly modules and live websites
  • Evidence-backed findings - explanations grounded in what it actually found, not guesses
  • Everywhere you work - in the UI, from the command line, or as part of your automations

🛠️ Improvements & Fixes​

Reliability​

  • Smoother remote sessions - large responses no longer break over the relay, and device listings load without errors
  • Clearer API errors - when something goes wrong, your automations now receive the original error details intact
  • Friendlier delegation - when an agent delegates work to another, you now see the agent's name instead of a cryptic identifier

TF Code​

  • Faster project loading - only your most recent projects load up front, keeping the interface snappy
  • Steadier sessions - session turns, agent syncing and voice recording all behave more reliably

Mobile​

  • Fresh store builds - version 1.0.9 of the mobile app is rolling out to the App Store and Google Play, with scheduling fixes and live voice improvements included

📋 Summary​

This update includes:

  • TF K3 Optimised - frontier intelligence with roughly 40% shorter thinking traces
  • Qwen 3.8 2.4T A95B - a long-horizon specialist for multi-day coding and research
  • RLVR tuning - train agents that check their own work with task-and-checker documents
  • Workspace model governance - admins control the model catalogue across app, TF Code and API
  • Mobile workspace - kanban boards, channels, routines and one-tap device peering
  • MCP media tools - voices, images and video from your automations
  • TF Code reverse engineer - understand any shipped software with evidence

Thank you for using ToothFairy! We're constantly working to improve your experience. If you have any feedback or encounter any issues, please don't hesitate to reach out to our support team at support@toothfairyai.com.

Keep building magic with AI! 🧚‍♀️✨

v0.708.0