Agent LMD · the model
LMD-4B: a local decision model
LMD-4B answers specific questions about a case: it picks one of the options it is given and states how confident it is. Agent LMD runs on those decisions, doing the clicking and the form filling.
What a decision looks like
The agent asks the model a question with a list of possible answers. The model returns a probability for each option. Where confidence is low, the agent hands the case to a person. High confidence is not a guarantee of a correct answer, which is why payments and important cases are confirmed by a person.
An agent decision
Which tool should be used?
What LMD-4B is not
- It is not a chatbot. It holds no conversation and answers no open questions. It answers only the questions the agent asks it.
- It does not generate text. Chat models such as ChatGPT write an answer word by word (autoregressively). LMD-4B reads the case and scores the options the agent offers in a single pass, returning a probability for each. That is why it is fast and runs on an ordinary graphics card.
- It does not invent answers. It can only pick one of the options given, a date or an amount found in the mail for instance. It will not add a drug or a date that is not in the case. It can still pick the wrong option, sometimes with high confidence, which is why the rules are enforced by the program and the important steps are confirmed by a person.
- It does not write mails. Where text is needed, a draft reply for instance, the program assembles it from a template and the values the model selected.
- It does not work alone. The agent decides which questions to ask and what to click, the model picks the answer, and you decide what needs a person's approval.
Underneath, LMD-4B is a transformer language model. We use it only to score options, never to generate text.
Results
| Test | LMD-4B (on your hardware) | Jev 1.13 (cloud) |
|---|---|---|
| JevBench-PL, a public test of decisions in Polish | 83.1% | 85.3% |
| JevBench-PL, hard questions | 69.4% | 72.1% |
| Adverse drug event reports (agent, 10 cases) | 9/10 | 5/10 |
Measured 29 September 2026. How we measured the agent tests
Hardware
The model runs on a single graphics card of the RTX 3080 class, or on an Apple-silicon Mac with at least 16 GB of memory. It needs no server room and no connection to a cloud. The exact requirements for your process are settled during the pilot.
Limits
- Confidence is not a guarantee: the model can be wrong even when it is confident.
- One question can carry up to 26 answer options.
- On the hardest, multi-step questions it does worse than the best cloud model.
- It prepares medical and legal decisions for a specialist. It does not take such decisions itself.
- The model alone is not enough: the questions we ask it, and the connection to your systems, are written for one specific process.