Voxta v1.9.3: Thinking Model Budgets, Adaptive Voice Detection & Agentic Tool Calling

Hello everyone,

This release is mostly about control and reliability. Local thinking models now have a proper reasoning budget, so you decide how much of the model's effort goes into thinking and how much is left for the reply. Voice activity detection was rebuilt to adapt to your room instead of relying on a fixed threshold. Claude Code gained a permissions system for tool calls, along with persistent memory and a native harness. Everything else is a long pass on the rough edges, and there were a fair number of them.

🧠 Local Thinking Models

Reasoning is now something you budget rather than something that happens to you.

• A reasoning budget for local thinking models, so you set how much effort goes into thinking

• Max thinking tokens are configured separately from reply tokens, so a long chain of thought no longer eats the reply

• Summarization services can use thinking too, instead of being limited to direct answers

• Qwen 3.8 formatting templates

📝 Prompting & Summarization

• The prompt now specifies the reply budget, so the AI can pace itself and avoid cutting off the important part

• Summarization handles edge cases, long messages and predictability better

• Use /title in a chat to generate a title for it immediately

• Affect prompting changed to reduce forced affects when they are not warranted

• H3 prompt improvements for multi-character scenes

🎙️ Adaptive Voice Detection

• New Adaptive Voice Activity Detection, built on NAudio 3, replacing the RMS Filter

• Deepgram Flux TTS support

• Deepgram SDK 7.0.0

• Fixed the listening window sometimes closing too early

• Fixed room noise delaying the end of a finished turn

• Audio capture improved and bandwidth reduced

🤖 Claude Code

• An agentic tool calling permissions system, so you approve what an agent is allowed to do

• Persistent memory across sessions

• A native harness

• Tool calls and tool permissions are rendered in the UI

• Tool calls are saved in the chat history

📦 Packaging & Export

• Chats export as voxpak and include generated images

• Voxpak dependencies support

• Warns when importing a voxpak with missing dependencies

⚙️ Models & Backends

• Updated llama.cpp (0.3.47) and ExLlamaV3 (1.4.2)

• Fixed the llama.cpp flash attention toggle not working

• Fixed llama.cpp receiving two BOS tokens on Llama 3 prompts

• Check for API key validity in OpenAI, XAI, OpenRouter and Mistral

• Anthropic refusals are reported instead of returning an empty reply

🔧 Key Fixes

A large share of this release went into fixing things that were quietly breaking.

• Fixed Whisper Live crashing on incomplete JSON for longer utterances

• Fixed Whisper Live hanging up after approximately 10 minutes without reconnecting

• Fixed infinite loops in scenarios crashing the whole process

• Fixed presets sometimes being overwritten silently

• Fixed Voxta Cloud dropping previous context when affects are used

• Fixed prosody increasing across follow-ups of the same affect

• Fixed a broken Models Directory preventing the module configuration pages from loading

• Fixed caching sometimes showing out of date resources after saving

• Fixed Linux encrypted package export corruption

🖥️ Voxta UI

• Search chats by title

• In the chats list, see how many messages each chat contains

• Reorder folders

• Each LoRA row is grouped into its own panel

• Preset modals and field ordering improvements

• See and change the current preset from module configuration

• Fixed models showing Not Downloaded when they are

• Fixed installing a new service and pressing Save and Install warning about unsaved changes

📱 Voxta Mobile

• Modals are now native bottom sheets with swipe to dismiss

• The stage view can display two participants

🆓 Voxta Lite

New to Voxta? Lite is free and runs on Voxta Cloud, so you can create an account and start chatting without any local setup. If you like it, a subscription gets you the full version, with a large library of integrations, custom modules, and the option to run everything on your own hardware.

Thank you all, as always, for the continued support and feedback. Whether you are tuning how long your model thinks before it answers, talking in a room that is no longer fighting you, or handing an agent a task with clear limits on what it can touch, we hope you enjoy this release.

Voxta is available to subscribers at: https://portal.voxta.ai/