Coming in 1.4: privacy you can see, safety you can prove: and a workspace that fits you

Valence has always made two quiet promises: your AI lives on your computer, and it's a space you can hand to the people you love. Version 1.4, in final testing as we write this, is the release where those promises stop being something you take on faith. Privacy you can see. Safety we can prove. A memory that listens, every model you bring working the way it should, and a workspace you can size to your own eyes. It's the steadiest, and most personal, Valence yet.

Private

Locked down

Encrypted on disk, with one switch that keeps everything on your machine.

Safe

Built for kids

Deny-by-default, real crisis support, and a red-team to back it up.

Smart

Memory that listens

Surfaces what's relevant, and your thumbs up and down actually count.

Flexible

Every model works

Web search, tools and reasoning behave the same, whoever you bring.

Yours

Sized to you

Scale the whole interface, chat, menus, all of it, to your screen.

Steady

Fewer rough edges

Crash fixes, no stray processes, calmer local answers.

Privacy you can see

Valence never needed an account and never sent your conversations anywhere. 1.4 makes that concrete:

Correction, 9 October 2026. This post said more than what shipped, here and in the "Private" pillar above. Encryption at rest is opt-in: it is off by default, and one switch in Settings (Data Encryption) turns it on for chat and memory files. Local-only mode blocks web search and the update check; it is not a promise that nothing can ever leave. The post is left as written. The current list of everything that goes online is on Privacy & Licensing.

Everything's encrypted on disk. Your conversations, your memory, and your profiles are now encrypted at rest, so even the files sitting on your own drive don't give up what you talked about.

One switch for "stays on this machine." A single local-only master toggle guarantees nothing leaves your computer, with plain-spoken labels beside every provider so you always know what would, and wouldn't, reach the internet.

Take it or erase it. Export everything you've saved in one step, or wipe it all just as easily. It's yours; you decide.

A space built for kids

Kids Mode grew up. In 1.4 it doesn’t just filter. It starts from no and opens only what’s been checked:

Deny-by-default. Every sensitive action is blocked first and allowed only when it's explicitly vetted. The safe direction to fail.

Help when it matters. If a child ever signals real distress, Valence surfaces age-appropriate crisis support instead of looking away, and quietly gives a parent a heads-up.

0

harmful outputs across 1,000 adversarial conversations in our own red-team. Not a promise. A result you can read for yourself.

Correction, 6 September 2026: this figure is withdrawn. The grader that produced it was tested for the first time in September 2026 and found to be marking every reply safe without reading it. What we found.

See the Kids Safety Proof →

Memory that actually listens

Valence’s on-device memory got noticeably sharper in 1.4: and it takes your feedback to heart:

Only what's relevant. Instead of pushing every saved note into every answer, Valence now weighs what you're actually asking and brings up only what fits, and stays quiet when nothing does. Cleaner answers, less clutter.

Your thumbs count. Thumbs-down a memory and it stops resurfacing; thumbs-up and it's favored next time. You're never stuck with what it guessed. You shape what it keeps in mind.

Works with every model

Valence has always let you bring your own AI, OpenAI, Claude, Gemini, xAI, OpenRouter, or a model running entirely on your machine. In 1.4 we went provider by provider and fixed the quiet places where the advanced features used to break, so switching models never costs you the good stuff:

Web search that actually searches. Ask Gemini something that needs the live web, "when does the Slate Truck come out?", and it searches, then answers. No error, no query that vanishes into an empty reply.

Deep thinking and tools, together. Claude's extended reasoning now rides alongside tool use the way it's meant to. The model can think hard and take action in a single turn.

Reasoning models, handled right. Point Valence at a step-by-step reasoning model and it speaks that model's dialect correctly, instead of stumbling on the way it likes to answer.

You're never locked into one provider to get web search, tools, or careful reasoning. Switch freely. The powerful parts behave the same no matter who's behind the curtain.

Size the whole thing to you

For a while, one setting changed your chat text and nothing else, so it never quite felt like it did anything. In 1.4 it became a real Interface Scale: it grows or shrinks the entire workspace, in five steps from Extra Small to Extra Large. Try it: tap through the sizes:

Valence

How should I plan a two-week trip?

Start with the anchors, the two or three places you can't miss, and let the days between them stay loose.

Ask anything…

Preview · the real thing scales the entire app, live

For every screen and every pair of eyes. Big, sharp monitor that makes everything feel tiny? Scale up. Want more on screen at once? Scale down. Whatever you choose, it's there again next time you open Valence.

Nothing gets left behind. Menus, dialogs, buttons and chat all grow together and stay where they belong. The whole interface reflows, so nothing clips off the edge or drifts out of reach.

Steadier under the hood

The unglamorous work that makes the rest trustworthy:

• Fixed a startup crash on certain graphics configurations that could leave the local model stuck.
• Stopped a rare out-of-memory crash on big, complex answers: heavy math and long code render safely now.
• No more stray background processes lingering after you close Valence.
• Calmer local conversations: the on-device AI no longer echoes your greeting back or leaves an awkward gap before it answers.

Almost here

None of this has shipped yet, 1.4 is in final testing as we write this, and we don’t put a version in your hands until it’s earned it. No firm date, but it’s close.

When it lands, it’s the same deal it’s always been: everything runs on your machine, no account required, and the 7-day trial becomes a one-time $59.95: never a subscription. If any of this matters to you, or there’s something we haven’t built yet, we’d genuinely like to hear it: help@helixailabs.com.

Now shipped. This is live in Valence 1.4.0.

Get Valence today →

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

← All news