Posted on Aug 1

TextRecast v0.2.0: User-Controlled Local AI Model Selection

#dotnet #ai #opensource #windows

I’ve released TextRecast v0.2.0, a major update to my local-first Windows writing assistant.

TextRecast lets you select text in another Windows application and improve it, shorten it, expand it, summarize it, or change its tone using a small language model running locally on your computer.

There is no account, API key, Ollama installation, Python environment, or cloud text-processing service involved.

What changed in v0.2.0?

The main addition is explicit, user-controlled model selection.

Before TextRecast makes any model-download request, it now displays the available models along with:

Their intended role
Download size
Hardware compatibility
English-language status
Known limitations
Whether the model is experimental

TextRecast downloads only the exact model the user confirms.

Available models

Role	Model	Download	Status
Fast/default	Qwen 2.5 1.5B Q4_K_M	1.04 GiB	Established default
Balanced	Qwen 3.5 2B Q5_K_M	1.34 GiB	Experimental
Best quality	Qwen 3.5 4B Q5_K_M	2.93 GiB	Experimental
Alternative	Granite 4.1 3B Q5_K_M	2.27 GiB	Experimental

The smaller Qwen 2.5 model remains the fast default. The additional models are clearly marked as experimental because local language models can still alter facts, roles, conditions, or deadline wording.

All current profiles are intended for English rewriting.

Model-specific behavior

Different model families do not always respond well to the same prompt or sampling configuration.

For this release, I added separate adapters for:

Qwen 2.5
Qwen 3.5
Granite 4.1

Each adapter can define its own prompt format, chat template, stop sequences, and sampling profile while sharing the same installation and inference infrastructure.

This keeps the application architecture extensible without requiring a separate model server.

Safer model downloads

The download process remains user-controlled and verifiable:

TextRecast shows the available choices.
The user selects a model.
The exact model and size are shown for confirmation.
Only the confirmed model is downloaded.
The final file must match its expected size and SHA-256 checksum.
Interrupted downloads can resume instead of restarting.

A valid installed model is reused on later launches.

Public release archives remain model-free, so users do not download several gigabytes of models they may never use.

Hardware-aware selection

TextRecast inspects available memory, CPU capabilities, and storage before recommending a model.

The recommendation is guidance rather than a hidden automatic decision. The user still sees the available choices and controls which compatible model is installed.

Only one model is loaded at a time.

Local Windows integration

TextRecast uses Windows UI Automation, clipboard fallback, and input APIs to work with selected text in other applications.

Before replacement, it verifies the source window, process, capture age, and selected content. The selected and generated text are not intentionally written to application logs or history.

Model inference runs locally on the CPU through LLamaSharp.

Testing and packaging

For v0.2.0, I expanded the English qualification corpus and added model-specific benchmark tooling covering:

Output quality
Semantic preservation
Operation intent
Latency
Time to first token
Generation throughput
Peak memory usage

The repository currently passes all 126 automated tests.

The Windows x64 build is self-contained, so users do not need to install .NET 10 separately. The release also includes the applicable project, dependency, icon, .NET runtime, and WPF legal notices.

The Microsoft Visual C++ x64 Redistributable may still be required.