Title
NOVA LocalAI Update: Voice, File Analysis and Optional VM Code Agent
Description
What's New
NOVA LocalAI has been upgraded into a portable, local-first AI workspace. It runs GGUF models directly through llama.cpp and does not require Ollama or an online AI service for normal chat, translation, file analysis, code generation, speech-to-text, or text-to-speech.
Three Fixed Local Models
The model selector now uses clear English names instead of long GGUF filenames.
Model
Best For
Qwen3 4B - Light Chat
Fast everyday chat, lightweight tasks, and quick responses.
Qwen2.5 7B - Translation
Chinese chat, long-text translation, document analysis, and multilingual tasks.
Qwen2.5 Coder 7B - Code
C++, Python, code generation, debugging, and project development.
The runtime uses the model's own GGUF chat template through the official llama.cpp Jinja template engine. This improves model switching stability without hard-coding a prompt format for a specific model name.
Modern English Interface
The interface, model labels, settings, status messages, and default conversation title are now English. Existing legacy default conversation titles are automatically migrated to New conversation.
File Attachments and Analysis
Code and text files can be attached as cards instead of being pasted into the input field. Each attachment has a remove button and can be used together with typed or spoken instructions.
Examples:
Plain Text
Explain this C++ file and find bugs.
Translate this document to German.
Summarize this file in Chinese.
Supported text-based formats include common C/C++ files, Python, JavaScript, TypeScript, JSON, Markdown, TXT, CSV, XML, YAML, logs, HTML/CSS, and DOCX paragraph text. Attachment tasks are processed in safe segments for long files.
Code Cards and Long-Code Protection
AI code blocks are displayed as dedicated code cards with Copy and Save actions. Direct code pasting now includes visible line and character limits. Normal code can be pasted directly, while oversized code is safely blocked and redirected to file attachment analysis instead of causing freezes or crashes.
Offline Voice Input and Windows Speech
Offline speech-to-text is supported through the optional whisper.cpp runtime. The microphone interface now shows recording status, animated waveform feedback, elapsed time, and transcription progress. Recognized speech is automatically returned to the input field.
Windows SAPI text-to-speech can read the latest non-code response. Speech settings support automatic reading, code-reading control, and switching between installed Windows voices.
Better Stop Behavior
Stopping a generation now stops future output while preserving already generated text as a Stopped response. Partial answers are no longer cleared from the conversation.
Optional Code Agent and Virtual Machine Support
A new Optional Code Agent is available in Settings. This feature is disabled by default and does not affect normal local chat usage.
When enabled, Code Agent automatically uses Qwen2.5 Coder 7B - Code. It can stage generated code into a user-configured shared workspace, request build or test commands, read build logs, and continue fixing errors from returned results.
The feature is designed for users who want to connect NOVA LocalAI to their own Windows development virtual machine. The application does not bundle a Windows image, a VM platform, or development tools.
VM Code Agent Workflow
1.
Install and configure your own VirtualBox or Hyper-V Windows development VM.
2.
Install the tools required by your projects inside the VM, such as Python, Git, CMake, Ninja, and Visual Studio Build Tools.
3.
Enable Code Agent in NOVA Settings and select the VM provider.
4.
Open the VM workspace from NOVA and map that workspace into the VM as a shared folder.
5.
Run the provided VM-side agent script in the shared workspace.
6.
Ask the Coder model to create C++ or Python code, then use Stage on a code card to place files into the shared workspace.
7.
When the model requests a build, test, dependency install, or web search operation, NOVA shows the exact action for approval.
8.
After approval, the VM performs the requested task and writes logs or results back to the shared workspace.
9.
Use Read latest VM build/test log in NOVA Settings to place the latest result into the chat input, then send it back to the model for analysis and fixes.
Optional Web Search
Web search is disabled by default. To use it, enable Web Search in NOVA Settings and allow network access in your own VM. The model can request a search term, but NOVA requires user approval before the VM performs the search. Search results are written to the workspace and can be returned to the model for research, documentation lookup, troubleshooting, or current-news tasks.
Safety Controls
NOVA does not automatically install packages, download dependencies, run unknown executables, delete files, or perform web searches. Those actions require explicit user confirmation. Normal local chat, translation, file analysis, and code generation remain available without configuring a VM or enabling network access.
Current Scope
PDF text extraction, legacy DOC reading, image understanding, multi-attachment analysis, and GPU acceleration are planned future improvements. The current model set is optimized for local text chat, translation, document analysis, and code tasks.