Voxta v1.9.2: Cloud Video Generation, Q8 Quantization & Sharper Memory (Patreon)

Hello everyone,

This one is a consolidation release. Since v1.9.0 we have shipped two updates, and this post covers both. The headline additions are video generation through OpenRouter and Voxta Cloud, Q8 quantization for llama.cpp, and a round of memory work that makes recall sharper and long chats cheaper to run. The rest is a broad pass on audio reliability, which had a few rough edges worth fixing properly.

🎬 Cloud Video Generation

Video generation is no longer tied to a local backend. OpenRouter and Voxta Cloud are both supported now, so you can generate without running anything on your own GPU. This is still experimental.

The playground only offers resolutions and durations the selected model actually supports

Reference to video, so you can drive a generation from an existing image

A polished viewer in the video playground

Videos download in the background on mobile

🧠 Sharper Memory

Memory got a focused pass aimed at recall quality and cost.

Memory recall now cuts noise and prioritizes keywords

The memory window is stabilized, which reduces prompt cache busting and keeps long chats cheaper

/memorize runs memory extraction immediately in a chat, instead of waiting for it to happen on its own

An Edit Memory button in the chat character's context menu

Fixed /summarize deadlocking

⚙️ Models & Backends

llama.cpp Q8 quantization support, now the default

OpenRouter model lists show pricing and feature information

Better rate limit handling

Temperature support removed from Anthropic, which no longer accepts it

ExLlamaV3 1.4.2 and ONNX Runtime 1.29.0

Resilient restore of built-in resources, with a warning if it fails

🔊 Audio & Voice

A large share of this release went into audio reliability.

New voice samples with affects (open Modules and use Install All to get them)

Prosody for narration is now carried separately from the character, so a narrator no longer inherits the character's delivery

Audio processing optimizations: silence is no longer sent over the network, and more of the work moved to the worklet

Fixed the audio connection keepalive being sent as audio, which produced 15ms of noise every 15 seconds

Fixed audio still being processed after turning off Speech To Text through the API mid-chat, along with several reliability improvements to audio connections and disconnections

Fixed "No module named 'pyaudioop'" for F5 TTS and OmniVoice

Fixed the Orpheus llama.cpp sampler against the new penalties signature

Fixed clip transcription returning "no result for this file"

🖥️ Voxta UI

Lists every release you do not have, not only the most recent one

Limits how many thumbnails generate at once on the characters screen

Retry a failed sign in, rather than leaving the button grayed out

Fixed images attached to assistant messages breaking OpenRouter and Voxta Cloud

📱 Voxta Mobile & Desktop

The mobile app draws edge to edge

Fixed the desktop app freezing with high RAM usage when an audio device cannot be found

Fixed a silent TTS breaking the diagnostics page

Stopped a double dispose of the IPC router from faulting shutdown

🧩 Voxta SDK

voxta-client is published against the SignalR version it is built with

🆓 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 generating video without a local GPU, fitting a larger model with Q8, or just enjoying a quieter audio pipeline, we hope you enjoy this release.

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