19th — train your own small task-specific models

TL;DR

Landing page for “19th,” a beta product that takes user-supplied examples for a focused task and automatically searches across architectures, sizes, and training strategies, then distills the best candidate into a deployment model. The headline pitch is a small on-device PII identifier and a Banking 77 classifier — the latter advertised as 160× faster, 310× cheaper, and +23.4 accuracy points over Opus 4.6 / Haiku 4.5. No technical writeup, no model details, no published benchmarks beyond the one chart on the page; treat as a product announcement, not a research artifact.

Key claims

19th lets users train task-specific models from scratch from labeled examples, with the explicit framing “ML team optional” [§landing page].
Two demo workloads are shown: a small on-device PII identifier (for pre-upload redaction) and a Banking 77 intent classifier [§landing page].
On Banking 77, the produced model is claimed at 12.6 ms vs. 2016 ms for Opus 4.6 (≈160×) [§landing page].
On Banking 77, the produced model is claimed at $0.03 / 1k tasks vs.$ 9.25 for Opus 4.6 (≈310×) [§landing page].
On Banking 77, the produced model is claimed at 94.6% accuracy vs. 77.5% for Haiku 4.5 (+23.4 percentage points after rounding) — note the comparison switches from Opus on speed/cost to Haiku on accuracy [§landing page].
The pipeline searches “hundreds of model candidates” across architectures, sizes, and training strategies, with “each generation informed by the last” — i.e. some form of iterative NAS-like search [§landing page].
Top-performing candidates teach smaller, faster deployment models — i.e. distillation into a production student [§landing page].
Outputs are gated as production-ready by passing robustness, latency, and cost thresholds [§landing page].

Method

No technical details are disclosed on the page. The product workflow described is: (1) user supplies examples, (2) 19th runs an iterative search over a candidate space of architectures × sizes × training strategies, (3) the best-performing candidates serve as teachers, distilling into a smaller student that meets latency and cost thresholds. No information on the search algorithm, the candidate pool, the distillation objective, the training data scale required from the user, supported task types beyond classification + PII, or whether models are produced on-device or via a hosted training service.

Results

The single quantitative comparison on the page is on Banking 77, a 77-intent banking-domain intent-classification benchmark. The asymmetry of the comparison (Opus 4.6 for speed and cost, Haiku 4.5 for accuracy) is worth flagging: the cheap-and-fast number is most favorable against the most expensive frontier model, and the accuracy number is most favorable against the cheapest one, so the chart cannot be summarized as a single “vs. Anthropic frontier” result. No latency / cost / accuracy comparison against e.g. a fine-tuned BERT, DistilBERT, ModernBERT, or other small-model baselines is shown, which is the comparison that would actually establish whether the search-and-distill pipeline adds value over standard recipes a practitioner could run themselves.

Why it’s interesting

The product is a direct commercial bet on the position On the Slow Death of Scaling argues for: that for many production tasks, a small task-specific model dominates a frontier LLM on the cost / latency / accuracy triangle, and the binding constraint is the engineering labor to train it — not the model itself. The Hooker essay names data curation, algorithmic technique, and distillation as the three levers that compensate for parameter count; 19th’s pitch is essentially “automate those three levers for one narrow task at a time.” Whether this is interesting as research depends entirely on what the automated search is doing internally, which the page does not disclose.

The page does fit one trend that’s filed elsewhere: AutoML/agentic-training systems (e.g. Introducing PostTrainBench (Thoughtful Lab) / PostTrainBench: Can LLM Agents Automate LLM Post-Training? benchmarking LLM-agents at automating post-training, Scaling Karpathy's Autoresearch: What Happens When the Agent Gets a GPU Cluster on agentic autoresearch). 19th is the productized end of that arc — a UI on top of automated training — but contributes no public technical content the wiki can cite.