Lor-1 v1.0 is GA

Lor-1 v1.0 is generally available. It is the in-house clinical model behind the Lorraine app and Lorraine Chat, and it is available to approved teams through the Lorraine Platform API. This post is a short tour of what we built and how we evaluated it.

What Lor-1 is

Lor-1 is an open-weight base model adapted in-house for South African clinical practice, quantised to INT4 before serving, and run on our own H200 capacity with speculative decoding to close the latency gap that quantisation and model size otherwise open. Specific hyperparameters, dataset composition, and pipeline internals are held privately and shared with qualifying enterprise customers under NDA, so this is not the place for a spec sheet. What matters at the bedside is the part you can feel: Chat answers at conversational speed, not batch speed.

Benchmarks

Every candidate checkpoint is evaluated on LorBench, our internal benchmark for South African clinical reasoning, across multiple decoding configurations before it can be promoted. LorBench deliberately includes questions where the SA answer diverges from the international one, so it measures geographic specificity and not just medical knowledge. Rather than publish percentage deltas here, we describe where Lor-1 has moved the needle: the largest gains land on SA drug dosing, and protocol retrieval remains the weakest of the five categories. The model card sets out what LorBench tests and how the promotion gate works.

What we learned

A few decisions earned their keep. Smaller, more conservative adaptations of the base beat aggressive ones — the aggressive runs looked better on the target benchmark for a while, then quietly forgot clinical knowledge the base already had right. Giving the model open access to its tools beat funnelling queries down a narrower path. Preference optimisation on top of supervised fine-tuning regressed three times in a row, so we stopped. And self-correction prompts consistently made the first answer worse, which is why Lor-1 is built to be right out of the gate rather than right on the second pass.

What comes next

v1.1 will focus on the two weakest question types — protocol retrieval and referral triage — on grounding failures where tool output was available but unused, and on a re-run of the training mix against an improved dataset that landed after v1.0 was already in the oven. Expect a follow-up post when it ships.