LastBrowser

Lastbrowser: a Windows browser for people who live in tabs and terminals. - vertical tabs + Spaces, separate cookies per account - built-in terminal (Ctrl+`) - your own AI key or local Ollama - history stays on your PC Free beta, no account: lastbrowser.com

LastBrowser just passed 1000+ downloads. Thank you to everyone who tried it, shared feedback and helped shape it. This is only the beginning. lastbrowser.com

LastBrowser just passed 500+ downloads. Thank you to everyone who tried it, shared feedback and helped shape it. This is only the beginning. lastbrowser.com

Lab notes from Serpentis, our small code-patching model. Progress report, not a launch. 1) We fixed our measurement first. Our eval checked patches with regex, so "green" sometimes meant "looks right" and not "works". Now a patch has to really `git apply`, and a task only counts as solved when the tests pass. Solved ≠ applied. Some old wins disappeared, but now we trust the numbers. 2) Pretty diffs vs. correct diffs. minimal_patcher v4 and v5 (LoRA, trained on Modal A100s) often produced patches that applied cleanly, and solved 0 smoke tasks. They had learned what a good diff looks like, not how to fix the bug. 3) v6: Exact-Line SFT. New training format: LOCATE the exact lines, then EDIT only those. 350 examples, loss ~0.066, ~$6 of GPU. Still 0 solved on the 2-task smoke test. But on one task (flit) the correct fix was fully inside the prediction for the first time. It failed because of an over-edit: the model also turned a `continue` into a `pass`. Our overlap metric said 1.0 because it doesn't count extra lines. That's measurement lesson #2. Next: decoding/filtering against over-edits (LOCATE lines only, fewest edits), a 32-task eval, and a paper-scope decision by Nov 10. No release, no leaderboard claims. Small model, honest eval, small budget. Full post: lastbrowser.com/blog/david-v…

Neuer Artikel: Ein Benchmark der SKRIME-Testserver im Vergleich zu unserem Gameserver. Cid hat gemessen, die Ergebnisse stehen im Artikel.

David vs. Goliath, Teil 1: Schlagen acht kleine Gemma-4-E4B-Spezialisten mit eigenen LoRAs ein 31B-Modell bei gleichem Compute? Unser Kaggle-Projekt von Anfang an, ehrlich: noch nichts trainiert, das Log ist leer. Plan, Architektur, Regeln: lastbrowser.com/blog/david-v…

TechLab.HK took a close look at LastBrowser: Spaces that keep your logins apart, Sidekick that only reads a page when you ask it to, a built-in terminal with manual approval, and an open-source GPL-3.0 license. Thanks for the thorough write-up! techlab.hk/p/lastbrowser-ai-…

LastBrowser just passed 100+ downloads. Thank you to everyone who tried it, shared feedback and helped shape it. This is only the beginning. lastbrowser.com

Local models in the browser feel great — until you hit the trust boundary. Passwords? Medical portals? Your inbox? Which browsing task would you still NOT hand to a local model, even a good one? (Building Sidekick into LastBrowser, so genuinely curious.)

Builders & researchers: what do you actually do with tabs you're not ready to read yet — leave them open, bookmark them, or dump into a read-later app? Has that habit changed since you started using vertical tabs or tab groups?

Local models in the browser: match the model to the job. A 3–8B via Ollama is plenty for quick page summaries, rewrites and pulling key points out, and it stays snappy. Save the big models for heavier reasoning. Task-matched beats biggest-every-time. Sidekick runs either, local or your own key.

The most underrated feature in any AI browser helper isn't the action — it's the undo. When it fills the wrong field or submits too soon, one clear step back matters more than a hundred clever steps forward. What's your "please let me undo that" moment?

Honest question: which Ollama model do you actually keep for reading and summarizing long web articles on a normal laptop? Not the benchmark pick, the one still installed after a week of real use. What decides it — speed, RAM, or how faithful the summary is?

Quick tip if you're trying Sidekick in LastBrowser with a local model: start with the boring tasks. Point it at Ollama or your own API key and give it something small — clean up a paragraph, extract a short list from a page, reword a sentence. Local models handle contained jobs better than people expect, and you get a feel for latency and quality without committing to a whole workflow. Genuine question: what's the first browser task you'd hand a local model? I'm curious where other people draw the line.

Which browser AI tasks belong on a local model, and which on a cloud model? My split: Local: summarize the page you're on, rewrite a paragraph, pull names or dates out of a tab, quick rephrases. Small task, small context, done instantly. Cloud: heavy multi-step reasoning across many tabs, long document comparison, anything that needs deep world knowledge. The trade-off: local keeps your data on your device, is fast, and costs nothing per request. Cloud is smarter, but your page content leaves the machine and you pay per token. LastBrowser's Sidekick can run with local models via Ollama or your own API key, so you can choose per task. Where do you draw that line?

Morning builder question: when an AI assistant summarizes a long page, do you want it to show which passages it used — or just give you the answer? I keep going back and forth on this while building LastBrowser. Curious what you'd trust more.

Building LastBrowser v0.1.x: early Chromium-based browser for Windows and Linux (Linux preview), with vertical tabs and built-in Sidekick that runs with local models via Ollama or your own API key. What I'm testing this week is research triage. I open 30 tabs, lose the thread, and the AI lives in another app. So the experiment: tabs on the left, Sidekick on the right, one task at a time. Summarize this page, remind me why I opened the last six tabs, keep moving. Still rough in v0.1.x, and that's honest. Some pages slow Sidekick down, some flows need polish. Open question: when research gets messy, does side-by-side AI sharpen triage, or just split attention twice as thin? How do you keep 20+ tabs from turning into noise? lastbrowser.com

Observation: everyone's putting AI in the browser, but nobody agrees where it should actually live. Sidebar wins when you're mid-page and just need a quick summarize or cross-check — the page is the context, the AI is the annotation. Separate agent tab wins when the task runs long: comparing sources, pulling a thread across a mess of tabs, research you don't want to babysit. So the real question: what's the first everyday task you'd actually trust it with? Summarizing a dense PDF? Turning your pile of open tabs into something readable? Double-checking claims on a product page before you buy? LastBrowser is Chromium-based, in Windows + Linux preview, with vertical tabs and an AI Sidekick that runs on Ollama local models or your own API key — so sidebar vs. agent tab stays an open question. Curious where you land: sidebar or agent tab? And what's the first task you'd hand over?

Builder note from LastBrowser (v0.1.x, Windows + Linux preview). We gave Sidekick a choice: run a small model locally through Ollama, or bring your own API key. People ask why both. Honest answer. What small local models are actually good at inside a browser: - Summarizing the page you're on. Tight context, clear task, fast. - Rewriting and tone shifts on selected text. Low stakes, easy to eyeball, instant. - Extraction: pulling names, prices, dates, key points out of the current page into something usable. Where they still fall short: - Multi-step reasoning across many tabs or long documents. They lose the thread. - Tasks needing broad world knowledge or precise citations. - Long agentic chains and tool use. They improvise, and it shows. So we don't pretend one model wins everything. Local is private, works offline, and costs nothing per token — great for the everyday work above. A larger model through your own key is better when the task is genuinely hard. That's the whole reason we let you pick. Per task, not per product. lastbrowser.com