v1.0 alpha · open source · MIT

A browser that thinks
with you, not at you.

Drag a paragraph into chat to cite it. Ctrl-click tabs to make them one context. Splits behave like your editor. Local model when you want privacy, frontier model when you want quality.

Free out of the box No API key required Windows · macOS · Linux Open source · MIT
D
DuckDuckGo
W
Mixture of experts (ML)
A
Mixtral of Experts
H
Qwen3-1.7B
P
Long-context evals
G
qwen3 / README
Mixtral of Experts arxiv.org · 2401.04088 ⌘L
arxiv.org · /abs/2401.04088 · jan 2024

Mixtral of Experts

We introduce Mixtral 8×7B, a Sparse Mixture of Experts language model. Mixtral has the same architecture as Mistral 7B, with the difference that each layer is composed of 8 feedforward blocks (i.e. experts). For every token, at each layer, a router network selects two experts to process the current state and combine their outputs.

Even though each token only sees two experts, the selected experts can be different at each timestep. As a result, each token has access to 47B parameters but only uses 13B active parameters during inference.

Pricing · or lack thereof

Free out of the box. BYOK if you want to.

Sable ships with a fast AI model that runs entirely on your machine — no key, no subscription, no per-message bill. For frontier-quality answers, paste your Anthropic or OpenAI key and Sable streams from those instead. You'll be billed by them on your own account; Sable itself never charges you anything.

A
Anthropic (BYOK)
Pay direct· at cost

Paste your sk-ant-… key in Settings. Sable streams from Claude — Anthropic bills your account. Stored in your OS keychain.

  • Frontier quality, near-instant first token
  • Multimodal — drop images straight in
  • Sable adds zero markup
O
OpenAI (BYOK)
Pay direct· at cost

Paste your sk-… key in Settings. Sable streams from GPT — OpenAI bills your account. Stored in your OS keychain.

  • Frontier quality, near-instant first token
  • Multimodal — drop images straight in
  • Sable adds zero markup
Switch providers per-task. Conversation history, citations, dropped images — all of it travels with you.
Why Sable

Built around the way you actually use AI.

Not a Chrome extension. Not a wrapper. A real browser where every part — tabs, layout, URL bar, sidebar — was designed to work with the AI, not next to it.

A chat sidebar that reads what you browse

Drop text — instant citation. Drop an image — the AI sees it. Ctrl-click tabs and they become one shared context for whatever you ask next.

Drag-to-split, drag-to-group

Drag a tab to a pane edge to split your view, as deep as you want. Drag one tab onto another to group — they line up side-by-side automatically.

Smart new-tab page

Type what you want and hit Enter. Sable picks the right destination — even if you didn't type a URL. Backed by your history and a small on-device model.

Three AI providers, one is free

Built-in offline model is the default — no key, no bill. Add an Anthropic or OpenAI key when you want frontier-quality answers.

Themed personalities for each space

Lavender, mint, coral, amber, rose, sky, sage. Each space tints the whole window so it feels like a different room.

Privacy-first by default

API keys live in your OS's password store. History stays on your machine. The on-device model needs no internet at all.

In practice

Here's what it feels like.

Four recipes for the kind of work Sable was built for. Each takes about 30 seconds of setup.

01 · Research

Compare three articles, find the disagreement.

"What do these three sources actually disagree on?"
  1. Open three articles in three tabs.
  2. Drag one tab onto another to group them — Sable auto-arranges them side-by-side. Drag a third onto the right edge for a 3-way split.
  3. Ctrl-click all three tabs to add them as chat context.
  4. Ask: "Compare the framing across these three sources. Quote where they disagree."
A
Source A
W
Source B
P
Source C
@AA @WB @PC
Compare these three sources. Quote where they disagree.
claude · sonnet 4
Where they disagree:
"…the labour-market effect was modest and short-lived…" Source A · ft.com · paragraph 4
"…unemployment among new grads spiked materially in Q3…" Source C · noahpinion · figure 2
A and B agree on policy framing; A and C diverge on the size of the effect…
02 · Code review

Diff and docs, side by side, asked together.

"Does this PR match what the docs actually promise?"
  1. Open the GitHub PR with the diff in the left pane.
  2. Drag the docs tab onto the right edge of the layout to split vertically.
  3. Each pane gets its own URL bar — navigate independently.
  4. Ctrl-click both tabs and ask: "Does this change match the right-pane spec? Any inconsistencies?"
G
PR #4218 · diff
D
docs / cache.md
@@ -41,7 +41,9 @@
- return cache.get(key)
+ const value = cache.get(key)
+ if (!value) return fetcher(key)
+ return value
apps/server/cache.ts
Cache layer · Behavior

When a key is missing, the cache must call the fetcher and back-fill silently. Callers always receive a value.

docs/cache.md · §2.1
03 · Trip planning

One space. One theme. Everything you need.

"Plan a long weekend in Tokyo from May 30."
  1. Make a new space called Tokyo; give it the Sky theme.
  2. On the new-tab page, type flights sfo to nrt may 30 — Google Flights opens with the right query.
  3. Open weather, hotels, neighborhood guides as additional tabs.
  4. Drag the four most relevant tabs together to group them.
  5. Ctrl-click the group and ask: "Build a 4-day itinerary across these tabs."
F
flights
N
neighborhoods
H
hotels
W
weather
space · tokyo · sky theme

Day 1 — arrive Narita, settle in Shibuya.

Crossing on foot to acclimatize. Light dinner at a standing soba bar near the station. Forecast looks clear after 6pm — perfect for a slow walk to Hachiko.
1Tokyo Cheapo · neighborhoods 2Tomorrow's Forecast · jma
04 · Offline

Works on a plane. Or anywhere else.

"I'm on a long flight with no Wi-Fi."
  1. Before takeoff, make sure the built-in offline model is downloaded in Settings (~1.1 GB, one time).
  2. In the air, switch the active provider to it.
  3. Chat works exactly the same. Citations, dropped passages, multi-tab context — all of it.
$0/forever No outbound network Apache 2.0
Active model
Qwen 3 · 1.7B
on-device · ✈ no network
ready
Personalities

Each space is a personality.

Spaces aren't just folders — each one re-tints the whole chrome with a pastel theme. The window genuinely feels like a different room as you switch.

Lavender
Mint
Coral
Amber
Rose
Sky
Sage
Roadmap

What's next.

A deliberately ambitious roadmap. Tier 2 and Tier 3 are where we'd most love contributors — open a discussion and pick something up.

Tier 1 · in flight

V1.0 polish — landing the alpha cleanly.
  • Tab grouping persistence across restarts.
  • Per-pane URL-bar drop targets — drag a tab onto another pane's mini URL bar.
  • Drag-to-rearrange for grouped tabs.
  • Light/dark auto-follow OS.
  • Code signing, auto-update, crash reporting.

Tier 3 · exploration

Personal knowledge & agentic browsing — the endgame.
  • Personal Knowledge Graph from history — encrypted-at-rest, default-deny blocklist, /recall over everything you've read.
  • Active learning — Sable predicts the tabs you're about to open at 9am.
  • Agentic green-thread tabs — spawn a hidden tab with a goal, capability allowlist, human-in-the-loop.
  • Generative surfaces — sandboxed iframes rendering AI-synthesized UI from open tabs.
  • Multi-window & detach — drag a pane out into its own window.
Built in the open

MIT licensed. Looking for contributors.

Sable is a small, ambitious project. The features above are real and shipping; the roadmap is real and waiting. If any of it sounds like the browser you wish existed, we'd love your help making it.