Skip to main content

🧚 DeepSeek V4.1 Flash, voice and computer-use local agents, and a fine-tuning overhaul

Released on September 12, 2026

DeepSeek V4.1 Flash lands — a multimodal model that sees images and reads text across a 1M context — alongside two powerful new ways to use your local agents in TF Code: talk to them by voice and let them drive your computer. We have also overhauled the fine-tuning experience end to end, refined Mobile Peering, and made usage easier to track! 🚀

🤖 New Models​

DeepSeek V4.1 Flash​

DeepSeek V4.1 Flash is a multimodal Mixture-of-Experts model that natively processes images and text. With a 552B backbone that activates only 16B parameters during decode, it cuts its cache footprint to roughly a quarter of DeepSeek V4 Flash — bringing strong, cost-efficient reasoning to everyday agentic work. Available in standard and priority tiers.

Key Features:

  • Multimodal - sees images and reads text natively, across a 1M token context
  • Cost-efficient power - a 552B MoE that activates only 16B at decode, for fast, affordable agentic workloads
  • Ready for agents - a reasoner with tool calling, planner and browser support built in

GLM 5.3 Flash Priority​

The priority tier of GLM 5.3 Flash is now available, offering faster, higher-priority access for quick, lightweight tasks.

🎙️ Voice in TF Code​

Talk to your local agents​

You can now speak to your local agents instead of typing. Voice mode brings speech-to-text input and spoken text-to-speech replies to TF Code, so you can drive a session hands-free from the terminal, the desktop app, or the web.

Key Features:

  • Speak and be spoken to - talk to your local agent with speech-to-text input, and hear it reply with text-to-speech, including spoken callouts for milestones and blockers
  • Hands-free or push-to-talk - let it listen and auto-send, or hold to record and release to transcribe
  • Easy to set up - toggle voice with /voice, and run tfcode voice setup to install the tools and grant microphone access

🎯 Computer Use in TF Code​

Let your local agent drive your computer​

Computer use lets a local agent control your real machine — the actual cursor, keyboard and screen — so it can complete tasks in any app you use. The agent looks at a screenshot, then moves, clicks, types and scrolls one step at a time until the job is done. It is opt-in and off by default.

Key Features:

  • Real GUI control - your agent can move the mouse, click, drag, scroll, type and press keys to work inside your real applications
  • Safe by design - per-action permission prompts, an action budget to stop runaway loops, and a full audit log of everything it did
  • Cross-platform and opt-in - toggle it with /computer-use; works on macOS, Linux and Windows with a guided setup for OS permissions

🧠 A Better Fine-Tuning Experience​

Prepare, train, version and serve your models​

Fine-tuning has been rebuilt end to end, from your training data through to serving the finished model. Prepare datasets more cleanly, configure training with friendly controls or expert overrides, and version, roll back and power your fine-tuned models with flexible serving — across the app, the API and MCP.

Key Features:

  • Smarter data prep - import multiple files at once with per-file validation, mark hold-out validation splits, and bulk-delete training documents
  • Configure your way - visual sliders for every hyperparameter, plus an advanced JSON editor for expert overrides that take priority
  • Version and serve - snapshot fine-tuned models as immutable versions, roll back to any pinned version, and choose on-demand, scheduled or always-on serving with instant power control

🖥️ Mobile Peering, Refined​

A smoother way to reach your desktop from your phone​

Mobile Peering — our beta for driving a local agent on your desktop from your phone — gets a polished redesign. A clearer layout separates same-Wi-Fi and over-the-internet relay modes, with a guided three-step setup, live status indicators and a built-in explainer to walk you through pairing.

Key Features:

  • Clearer modes - distinct cards for same-Wi-Fi and over-the-internet relay, each with its own status
  • Guided relay setup - a three-step flow takes you from relay address to pairing code with feedback at every step
  • Friendlier names - the section is now labelled Phone access, with simpler wording throughout

Note: Mobile Peering remains in beta and will be made available soon to all users and regions.

🛠️ Improvements & Fixes​

Reliability and polish​

  • Code Execution costs - see Code Execution as its own line item in your consumption dashboard
  • Proxy scraping - let agents scrape web content through a proxy during internet search, for tougher sites
  • Better tool calling - smoother, more reliable tool calling for your cloud agents
  • Mobile app 1.0.7 - keyboard no longer hides the header on the peering and remote chat screens, plus general polish
  • Fine-tuned models in chat - fine-tuned models now appear correctly in the chat picker with the right pricing, with no silent fallback to a base model

📋 Summary​

This update includes:

  • DeepSeek V4.1 Flash - a multimodal, cost-efficient MoE model with 1M context, vision, tool calling and reasoning, in standard and priority tiers
  • GLM 5.3 Flash Priority - the priority tier of GLM 5.3 Flash for quick, lightweight tasks
  • Voice in TF Code - talk to your local agents with speech-to-text input and text-to-speech replies
  • Computer Use in TF Code - let a local agent control your real mouse, keyboard and screen, safely and opt-in
  • Fine-tuning overhaul - smarter data prep, visual or JSON config, and adapter versioning, rollback and flexible serving
  • Mobile Peering - a refined, guided beta for driving a local agent from your phone
  • Reliability - Code Execution costs, proxy scraping, better tool calling, a smoother mobile app and correct fine-tuned model pricing

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! 🧚‍♀️✨

v0.705.0

Last updated on