It trains between requests
It trains a LoRA adapter on the base model the engine already holds in memory, using the idle time between requests, and then serves that adapter from the same machine.
Read the paperuli fine-tunes a small model on your own labelled data, inside a running inference engine on a Mac, in the idle time between requests.
Free. Sign-ups close 20 November. How it works for testers
$ uli submit train.chat.jsonl --base qwen3.5-9b --name banking77-20261029-184325 dataset 2,000 examples · 77 labels · 0 bad rows · 724 KB uploaded sha256 27c2…76f6 assigned node e2e-node · qwen3.5-9b resident · idle job banking77-20261029-184325 (j-bc722f26) queued → training $ uli watch banking77-20261029-184325 banking77-20261029-184325 · qwen3.5-9b · rank 8 · 3 epochs training progress █████████████████████████████░ 100% update 1,499 / 1,500 epoch 3 of 3 loss 2.17 █▃▃▃▂▂▂▂▁▁▁▁ 0.00 elapsed 8h 29m eta ~0h checkpoint update 1475 · 7 min ago node e2e-node · serving 0.0 req/s alongside training gaps 90% idle → training · longest piece 0.09 s ✓ banking77-20261029-184325 succeeded in 8h 29m adapter qwen3.5-9b:banking77-20261029-184325 · 62 MB · stored on hub
The real output of the overnight run of 29 to 30 October 2026. This overnight run served no chat traffic, so it trained the whole time.
Read the paperA labelled dataset of a few hundred rows or more, one example per line.
Getting startedWhen a customer writes My card has not arrived, the answer is card_not_arriving.
uli checks every row before uploading and lists any line that breaks the format.
Grade your model, then the base model given every label, on the same held-out rows.
It trains a LoRA adapter on the base model the engine already holds in memory, using the idle time between requests, and then serves that adapter from the same machine.
Read the paperOn a 77-intent banking task, the trained adapter answered 90 of 100 held-out questions from the question alone. The same base model, given every label, answered 69.
See the resultsThe whole loop, from training to serving, runs on a 16 GB M1 Pro.
In the serving test, training added at most 0.134 s to a waiting request (p99, at 0.2 requests per second).
Killed mid-job, it resumed from its last checkpoint and finished with an identical adapter. The training data is deleted from the Mac when the job ends.
Serving a generic model is a race on price and speed, and large data centers win it. A model trained on one customer's labels is a different kind of work. Its value comes from that customer's data, not from scale, so it suits a Mac that already holds the base model and has time to spare.
uli runs inside EngineV2, the open-source inference engine for Apple Silicon that Darkbloom runs on its providers' Macs, published under the MIT license in Layr-Labs/mlx-swift-lm. uli adds an idle hook and the training loop. It is an independent research project, not an official Darkbloom product, and it uses no code from Darkbloom's provider app.
An ongoing research project, measured so far on one Mac and one base model, Qwen3.5-9B in 4-bit. The paper gives every number with its source and its limits.
Read the paperThe first cohort is free and starts with the Darkbloom community. Testers train a model on their own data and use it, on a Mac uli runs. Getting started shows every step, and how your data is handled says what is stored and when it is deleted.
Sign-ups close 20 November.