Valence 1.9.0 - Watch it think, stop it any time

Thinking you can follow, a Stop button that always works, and stronger Kids protections.

What’s new

• See what the model is thinking. While a local model thinks, a single line under “Thinking” shows what it is thinking about right now. Click it to open the full thought, which stays in a small box that follows the newest line.
• Tool steps show as steps. When the model looks something up or does a sum, each step appears as its own line above the reply, never as code in the answer. If it uses several tools in a row, every step’s text stays on screen and in the saved chat.
• Long chats keep going. When a chat with a local model gets too long for the model’s memory, the oldest messages slide out instead of the reply failing. A line in the chat shows where the model’s memory now starts.
• Stronger Kids protections. Text, attachments and open tabs from a Kids profile never carry over to an adult profile, or the other way round. Claude and Gemini always keep the Kids safety instructions, even in long chats. When a parent continues a Kids chat on a cloud model, Valence asks first.
• Images in replies load only when you click. Pictures linked in a reply no longer load by themselves, and in Kids profiles they never load.
• Voice typing is paused. Earlier versions sent your voice to Microsoft to turn it into text. Valence no longer does that. The recognizer built into Windows works offline but isn’t accurate enough yet, so voice typing is off in this version.

Fixes

• Stop works from the first second. The Stop button appears as soon as you send. Stopping while the model is thinking keeps the reply as “Thought for 6s”, and stopping no longer shows “The operation was canceled.”
• Scroll up and it stays put. If you scroll up while a reply is being written, the chat no longer pulls you back down. Click the arrow at the bottom to catch up.
• The local AI restarts itself. If the local model freezes or quits, for example after sleep or a screen lock, Valence restarts it quietly. You only see a message if the restart fails.
• Fewer empty replies. If a thinking model finishes without an answer, Valence asks again once with thinking off.
• Smaller things. Settings tabs renamed (Contexts is now Memory Spaces, Context is now Long Chats); a new Neural Net loading animation; pop-up messages are solid and easier to read; the Qwen 27B IQ2 model shows “In testing” in the model list.

1.9.0 also carries everything in 1.8.3.

Download

Valence 1.9.0 — Slim (.msi) · 122 MB · your GPU pack is fetched on first run.

Valence 1.9.0 — All-in-one (.msi) · 620 MB · CUDA + Vulkan bundled, fully offline.

Windows 10/11 (64-bit) · 7-day free trial, then $24.99 for Windows, or $59.95 for every platform.

All-in-one SHA-256: 58F14A19E186DEC82CD171EF150042CC4F219E83B22D201FBACC36818923CCEC Slim SHA-256: 8AA90809C548A31A21C81B9887C37B5429F7B56BF42A8B6E5105E377E2559517 Unsigned installer — SmartScreen may warn: More info → Run anyway. Existing installs upgrade in place.

← All updates