Most of the AI that helps you code lives in someone else's cloud. Your code goes up, the answer comes back down, and the whole arrangement quietly assumes you'll always have a connection, a subscription, and a willingness to let your work leave your machine.
169Pi launched something today that assumes none of that. Alpie Seed is a 9-billion-parameter open-weight model — Apache 2.0, so you can actually use it commercially — that codes, reasons, and understands text, images and video. And it runs fully offline on hardware you probably already own: a regular laptop, a phone, even a Windows PC with no graphics card at all.
That's the whole story, really. Not the benchmark table — the fact that the model is small enough to live where your code lives.
What Alpie Seed actually is
The specs, as published by 169Pi on 6 November 2026:
- 9 billion parameters, open-weight under Apache 2.0 — free to use, modify and ship, including commercially
- Multimodal: text, images and video, not just code and chat
- Trained in 4-bit from day one — the team says that was the only way a small team could afford to build it, and it's also why the model is so small to run
- About 5 GB — that's the number that matters most. A 5 GB model fits on a laptop the way a 70 GB one never will
The company behind it is small and, frankly, mostly unknown — this is not a household-name lab. That matters for how much weight to put on the claims below, and I'll keep flagging it. But the shape of the launch is what caught my attention: a model designed from the start to run on the edge, not shrunk down to fit there afterwards.
The vendor numbers, in one table
169Pi published benchmark results with the launch. Every number here is vendor-reported — the technical report, benchmark repository and testing instructions are public, so anyone can run the tests themselves, but nobody independent has yet. Read the table with that in mind:
| Benchmark | Alpie Seed (vendor-reported) | Comparison |
|---|---|---|
| SWE-bench Verified (fixing real issues in real GitHub projects) | 61.2% | Qwen3.5-9B 60%, Sarvam 30B 34% |
| LiveCodeBench v6 | 72.99% | Qwen3.5-9B 65.6%, Sarvam 30B 70% |
Two things worth noticing. First, the comparison the vendor chose is pointed: Qwen3.5-9B is the obvious same-size rival, and Sarvam 30B is a much bigger model that scores far lower on SWE-bench Verified — 169Pi is making a "small beats big" argument. Second, these are the vendor's own test runs, selected for a launch announcement. The numbers are plausible for a well-built 9B model in late 2026, but plausible is not verified. If you care about the benchmarks, the honest move is to run the public test suite yourself — 169Pi says you can.
What "runs on your laptop" actually means
The speeds 169Pi reports, per device:
| Hardware | Reported speed |
|---|---|
| RTX 3050 laptop GPU | ~20 tokens/sec |
| MacBook M2, 16 GB | ~14 tokens/sec |
| iPhone 17 Pro | ~9.6 tokens/sec |
Those numbers are usable, not blazing. 20 tokens a second is a comfortable reading pace for generated code; 9.6 on a phone is slow enough that you'd feel it on long outputs, fast enough for short answers and edits. The point isn't that this replaces a frontier cloud model on raw speed. The point is that it works at all — on a flight, in a village with patchy signal, inside an office that can't let data leave the building, on a Windows PC with no GPU. The model answers faster than you can read, in the vendor's words, and on a laptop that's roughly true.
Where to get it today
Alpie Seed is available now on Hugging Face (169Pi/Alpie_seed), Ollama, LM Studio and the PocketPal app, and you can try it without installing anything in the Alpie playground at playground.169pi.ai. The full model page is at 169pi.ai/research/alpie-seed-9b.
Alpie Local: the coding agent built on top
The second half of the launch is Alpie Local — a coding agent built on Alpie Seed that runs on your own machine. It plans work, reads your codebase, edits files, fixes bugs, runs tests and self-corrects, and it works with images, videos and GitHub repos. It's invite-only from today, with wider access later this month; the waitlist is at 169pi.ai/alpie-local.
A VS Code extension and fully offline enterprise deployments are planned next. The extension is the one to watch — a local coding agent is only as useful as the editor it lives in.
What this actually changes for you
If you code where the internet is unreliable — flights, trains, rural areas, conference Wi-Fi — this is the first genuinely capable open model that fits that life. A 5 GB download once, and your coding assistant works everywhere your laptop does. No per-token bill, no connection required.
If your code can't leave the building — client work under NDA, regulated industries, a company policy that says no cloud AI — the argument writes itself. The model runs where the data already is. Nothing is sent anywhere, because there's nowhere for it to be sent.
If you just use cloud coding agents, nothing changes tomorrow. A 9B local model won't out-code the frontier models you already pay for, and at ~14–20 tokens a second it won't outrun them either. What changes is the fallback: when the connection drops or the API bill stings, you now have a real local option instead of nothing.
Try it yourself in ten minutes
- Pick your runner. Ollama or LM Studio if you already use one; PocketPal if you want it on your phone; the Hugging Face page if you want the raw weights.
- Give it a real bug, not a hello-world. The honest test for a coding model is a bug in your own repo — something with context, not a LeetCode prompt. SWE-bench is a bug-fixing benchmark for a reason.
- Watch the speed on your hardware. The vendor's tokens-per-second numbers are for their test setups. Yours will differ. If it feels too slow for your workflow, you'll know in five minutes.
- Try an image or a screenshot. The multimodal claim is the least-proven part of the launch — a screenshot of a broken UI, a photo of a whiteboard sketch. See what it actually does with it before you believe the brochure.
- Compare it against what you already use. Same prompt, same bug, your cloud assistant and Alpie Seed side by side. That's the only comparison that matters for your work.
The limits — stated plainly
- All benchmark numbers are vendor-reported. The test suite is public and reproducible, which is good — but until someone independent runs it, treat 61.2% and 72.99% as claims, not facts.
- 169Pi is a small, little-known startup. There's no track record to lean on here, no big-lab reputation behind the numbers. The Apache 2.0 licence and public benchmarks are the trust mechanism, not the brand.
- 9B is 9B. This model will not match frontier cloud models on hard reasoning or long, complex refactors. Its advantage is where it runs, not how smart it is.
- Phone speeds are slow. ~9.6 tokens a second on an iPhone 17 Pro is usable for short work and painful for long generations. The phone story is real but modest.
- Alpie Local is invite-only. The coding agent — arguably the more interesting half of the launch — isn't something you can just download today. The waitlist is the only door in.
- Nothing here is hands-on. I haven't run Alpie Seed myself yet. No independent benchmarks, no long-term reliability data, no word on how it holds up over a full working day.
The bottom line is a small, honest one. Alpie Seed doesn't try to be the smartest model of 2026. It tries to be the capable model that's actually there — on your laptop, on a plane, inside a locked-down office — when the cloud isn't. If the vendor's numbers hold up under independent testing, that's a genuinely useful thing to have, and it costs you nothing but 5 GB of disk. If they don't, it's still an Apache 2.0 model you can test yourself in ten minutes, and the test is the whole point.
Have you run a local model for real coding work yet — not as a demo, but for a full day? What broke first: the speed, the context, or the quality? I'd like to hear what your setup actually looks like.
Sources: CXO Today, "169Pi Launches Alpie Seed, an AI Model That Codes, Sees and Reasons on Your Laptop Without the Internet" (6 November 2026, published 09:34 UTC); Hugging Face model page https://huggingface.co/169Pi/Alpie_seed; 169Pi model page https://169pi.ai/research/alpie-seed-9b. All facts come from the launch-day reporting, checked 6 November 2026. Benchmark figures are vendor-reported, not independently verified. No hands-on testing was performed for this article. This breaking article is not the daily buyer-intent post, and it contains no affiliate links.
