Valence 1.6.0 - Smooth answers, and memory that shows its work

This release is mostly about one thing: the app getting out of your way. Answers arrive smoothly, they finish, and when your AI remembers something it shows you what it used.

We previewed this work here: the pause fix and the memory redesign.

What’s new

• The pause is gone. Replies used to stall partway and land in a lump. A stuck internal check could hold back the entire rest of an answer. Fixed, along with the layers under it: streaming no longer waits on disk writes, no longer re-measures the window sixty times a second, and no longer slows down as answers grow.
• Answers that actually finish. The reply budget now comes from your model’s real capacity instead of a fixed number, about double the old ceiling on a typical setup. When a limit genuinely is reached, Valence says so instead of trailing off mid-sentence.
• Local models run faster. GPU use is now sized to your hardware instead of forced, and Qwen 3.8 loads at last (updated bundled engine).
• Memory that answers the question. “Did we talk about X?” actually searches; delivered memories get used; paraphrases land; junk stops crowding out real notes; stranded memories from other profiles come back automatically. One “memory recalled” chip sits with the answer it informed.
• X-Ray shows the receipts. A new Performance card with the engine’s own measurements, and X-Ray no longer asserts things it cannot know.
• Kids Mode, wider crisis detection: indirect phrasing is recognised, the AI’s own replies are checked, and a model must pass our red-team to be offered at all. Re-audited for this release: 1,000 fresh adversarial turns per model against the exact shipping safety system, zero harmful replies reached the child. Read the full audit. Correction, 6 September 2026: that “zero harmful replies” figure is withdrawn — the grader that produced it was later tested and found to be marking every reply safe without reading it. What we found and what changed.

Fixes

• Hover buttons on messages are clickable again (including Kids read-aloud); hidden menu items no longer steal clicks; the tall gap under replies is gone.
• Local maths renders properly typeset; Repeat Penalty can no longer be set to a value that produces gibberish.
• X-Ray content stays inside its window; the context meter agrees with the model’s real window.
• Valence’s own icon finally appears in Alt-Tab and taskbar previews.

Download

Valence 1.6.0: All-in-one (.msi) · 808 MB · CUDA + Vulkan bundled, fully offline.

Valence 1.6.0: Slim (.msi) · 310 MB · your GPU pack is fetched on first run.

Windows 10/11 (64-bit) · 7-day free trial, then $59.95 once.

All-in-one SHA-256: 47E902E4F4B15A016535F7509DC5DB2DFDC6DFAB1B1B4C8BF269228CE39229A3 Slim SHA-256: 31F3C1E3A9A7BBF70B79B88840C2179E79D8232ED9675E6D34D58881E271CA65 Unsigned installer, SmartScreen may warn: More info → Run anyway. Existing installs upgrade in place.

← All updates