LALAL.AI put real-time stem separation inside a $349 guitar device, fully offline

LALAL.AI now runs inside a guitar pedal-sized box. At IBC2026 in Amsterdam, which ran September 11 to 14, 2026, the company showed its stem-separation engine running natively on Neko, a compact wireless studio built for guitar players by Neko Engineering. The device splits a track into vocals, drums, bass, and guitar in real time with no internet connection, because the model runs on a neural chip inside the box instead of on a server.

No upload means no wait, and no per-use cost for running the model in someone else’s data center.

“

For six years our tech has lived in a browser tab. Putting it inside a physical device, running in real time with nothing sent to a server, is a different kind of product decision.

What Neko actually does, and which part still needs Wi-Fi

Neko is a pocket rig rather than a plugin host. It carries a digital 5.8 GHz wireless guitar link with 6.3 mm in and out jacks, amp and effects modelling, onboard recording, AI arrangement, and audio-to-tab transcription. Battery life is quoted at 3 hours or more. Pre-orders run at $349 against a $499 retail price, with a Kickstarter planned for October 2026 and shipping to backers starting in 2027.

The separation is the offline part. Neko Engineering CEO Vadim Smirnov said the priority was stem separation a musician can use while playing, without waiting for a network. Tab transcription is the exception and still runs in the cloud, so a practice session on a train gets you isolated stems but not the tab.

Put that next to the rest of the hardware moving this way. JBL built stem separation into a practice amp earlier this year, and LALAL.AI shipped a VST plugin that separates offline inside a DAW in May. Separation is quietly becoming a feature of the gear rather than a website you paste a file into.

Why a model that runs locally changes what stem separation costs you

Cloud separation has a per-minute price, a queue, and an upload. For a bedroom producer pulling a drum stem out of a reference track, that is an annoyance. For anyone working on unreleased material, sending a client’s master to a server is the part that actually blocks the workflow, and it is the reason studios with confidentiality clauses keep doing this by hand.

Local processing takes the recurring cost out. You pay for the device once, and the 500th separation costs the same as the first, which is zero. That maths is what makes on-device models attractive to anyone running stem separation tools daily, and it is the same reason phone makers moved speech recognition on-device.

The catch is model size. A phone-class chip runs a smaller model than a GPU farm, and LALAL.AI’s cloud service now splits up to 10 stems, including the separate Guitars stem it added to its Andromeda model in August. Neko separates 4. Nobody has published a quality comparison between the on-device split and the cloud one yet, and that comparison is the thing worth waiting for before you swap a paid plan for a $349 box.