A domain-specific LLM is a large language model trained to know one field, its vocabulary, facts and conventions, so that it can serve many tasks inside that field. Finance, law, medicine and chip design each have published examples. A task-specific model is narrower: it is trained for one job, such as summarizing a meeting for a query or pulling fields from an invoice, and it is judged on that job alone.
The two ideas often share one definition. IBM's explainer, read on 29 October 2026, describes a domain-specific LLM as one "trained or fine-tuned to specialize in a specific field or subject area, allowing it to perform domain-specific tasks more accurately and efficiently than a general-purpose LLM." That sentence covers both. This post separates them, because the training data, the build and the evaluation differ for each, and most real systems use the two together.
A note on the search term: in software, a "domain-specific language" is a small programming language such as SQL, so results for "domain-specific language models" mix two subjects. This post is about LLMs.
What is a domain-specific LLM?
A domain-specific LLM is a language model whose knowledge has been shifted toward one field by training on that field's text. It reads the field's documents with fewer misreadings, uses its terms correctly and follows its conventions, such as citation formats in law or the structure of a clinical note. Inside the field it stays general: one finance model can answer questions about filings, classify news sentiment and draft a summary.
Three routes produce one, from the most expensive to the least:
- Training from scratch on a mix of domain and general text, as BloombergGPT did.
- Continued pretraining of an existing model on raw domain text, as SaulLM-7B and BioMistral did.
- Fine-tuning on domain instructions and examples, usually after continued pretraining.
Our post on continued pretraining vs fine-tuning vs RAG covers when each route fits, and when retrieval does the job with no training at all.
How is a domain-specific LLM different from a task-specific model?
A domain-specific LLM knows a field and serves many tasks in it, while a task-specific model does one job and is measured on that job. The difference shows up in four places.
| Domain-specific LLM | Task-specific model | |
|---|---|---|
| Scope | One field, many tasks | One job |
| Training data | Raw field text, then instructions | Inputs and outputs of the job |
| Typical method | Continued pretraining, then fine-tuning | Fine-tuning on task examples |
| Evaluated on | A suite of field benchmarks | A held-out set from the job |
| Example | A legal LLM | A contract-clause extractor |
The quickest test is the evaluation. If one held-out set can decide whether the model is good enough, you are building a task model. If you need a suite across many tasks, you are building a domain model.
Research drew the same line in 2020. In "Don't Stop Pretraining", Gururangan et al. found across four domains and eight classification tasks that "a second phase of pretraining in-domain (domain-adaptive pretraining) leads to performance gains", and that adapting to the task's own unlabeled data "improves performance even after domain-adaptive pretraining." Domain knowledge and task skill are separate gains, and the second builds on the first.
Our foundation models vs task-specific models explainer covers the task side in depth: what task-specific models are, when they win, and how a business's work can be split across specialists.
What are examples of domain-specific LLMs?
Published domain-specific LLMs cover finance, law, medicine and engineering. Each example below was read at its source on 29 October 2026.
| Model | Field | Route | Published |
|---|---|---|---|
| BloombergGPT | Finance | Trained from scratch | March 2023 |
| SaulLM-7B | Law | Continued pretraining, then instruction tuning | March 2024 |
| BioMistral | Medicine | Continued pretraining on PubMed Central | February 2024 |
| Cheng et al. models | Biomedicine, finance, law | Continued pretraining on reading-comprehension text | September 2023 |
| Cadence Maestro | Chip and system design | Domain post-training | 26 October 2026 |
The table shows every route in use, and the newest entry combines domain models with general ones.
BloombergGPT is a 50-billion-parameter model trained on a 363-billion-token financial dataset plus 345 billion tokens of general data. Its paper reports that it "outperforms existing models on financial tasks by significant margins without sacrificing performance on general LLM benchmarks."
SaulLM-7B starts from Mistral 7B and is trained on "an English legal corpus of over 30 billion tokens", then instruction fine-tuned on legal datasets, per its paper.
BioMistral is Mistral "further pre-trained on PubMed Central" and evaluated on 10 medical question-answering tasks, per its paper.
Cheng et al., in "Adapting Large Language Models to Domains via Reading Comprehension", found that continued pretraining on raw domain text "endows the model with domain knowledge, but drastically hurts its prompting ability for question answering." Turning that text into reading-comprehension exercises improved results across all three of their domains, and the authors report their 7B model as competitive with much larger domain models such as BloombergGPT-50B.
Cadence Maestro, announced on 26 October 2026, is a portfolio of domain-specialized post-trained models that Cadence's agents orchestrate alongside frontier and open-weight models. Cadence reports initial results of up to 30% higher quality and up to 5x greater token efficiency against a general-purpose foundation model, and says the first models reach select early-access customers in the fourth quarter of 2026.
How do domain-specific and task-specific models combine?
They combine in three ways, and production systems usually use more than one.
- Stacked training. Domain pretraining first, then task fine-tuning on top. SaulLM-7B follows this order, and the Gururangan et al. results explain why it works: each stage adds its own gain.
- Routing. A general or domain model takes open-ended requests, and task specialists take the jobs that repeat at volume. Cadence's design is a public example: agents decompose a goal and send each step to a domain model, a frontier model or an open-weight model.
- Retrieval. Facts that change go into a retrieval index, and behaviour, format and task skill go into the model. That split is the principle we use in every build.
Training a model on the inputs it will actually see is where the largest gain sat in our own research. In our QMSum study (arXiv 2609.25028), one base model reading retrieved transcript spans gained 1.55 ROUGE-1 on the test split from the better inputs alone, and a further 5.29 ROUGE-1 once it was trained on those same span-shaped inputs, with a 95% interval of +4.02 to +6.56. Retrieval and training worked best together, and training carried most of the gain.
When does a general model do as well?
A general model does as well when the field's knowledge is public and already well covered in its training data. Manufacturers' Monthly's comparison of domain-specific and general-purpose models (20 October 2026) notes that BloombergGPT's lead narrowed once newer general-purpose models arrived, and that a prompted general model later matched a medical fine-tune on licensing-exam questions.
That pattern points at where domain training keeps its value: knowledge the general models never saw. Your organisation's documents, internal vocabulary, decisions and formats sit outside every public training set, and so do the specific jobs your team runs every day. The same article lists the other reason teams build their own: data privacy and regulatory constraints that rule out externally hosted AI services.
| Situation | Better fit |
|---|---|
| Many tasks in one specialised field | Domain-specific LLM |
| One job repeated at volume | Task-specific model |
| Field knowledge that is public and well covered | Prompted general model |
| Facts that change every week | Retrieval |
| Private vocabulary the base model never saw | Continued pretraining or fine-tuning |
| Data that must stay in your infrastructure | A trained open model |
Most teams land on a mix: a capable model for the open-ended work, trained specialists for the repeated jobs, and retrieval for the facts.
What does a trained model do to cost?
A trained model, domain or task, lowers the bill on two fronts: a lower price per token than the large frontier models, and fewer tokens per task, since the instructions, glossary and context you repeat in every prompt get trained into the model itself. The second saving scales with how much of each prompt is repeated boilerplate, so measure that share on your own traffic before putting a number on it. The cost-per-task calculator works through both dimensions with your own prompt sizes and volumes.
Ertas is an applied AI lab, and the adaptive model endpoint is how we put a trained model into production: a model trained on your task and your field's data, served through an OpenAI-compatible API. New versions train on your agents' traffic in the background, are scored against your evaluations, and go live only with your sign-off. If a new version falls short, the current model stays in service. It runs Ertas-managed, in your own cloud or on-premise. The endpoint is currently invitation only, and we are developing a more streamlined version for wider access.
Frequently Asked Questions
Is ChatGPT a domain-specific LLM?
ChatGPT runs on general-purpose models trained on broad data. Given instructions and documents it can work inside a field through prompting and retrieval, while the model itself stays general. A domain-specific LLM has had its training shifted toward one field.
What is the difference between a domain-specific language and a domain-specific LLM?
A domain-specific language is a small programming language built for one area, such as SQL for databases. A domain-specific LLM is a large language model trained on one field's text. The two share a phrase and little else.
How do you build a domain-specific LLM?
Start with an existing open model, continue pretraining it on raw text from the field if the base model lacks the vocabulary, then fine-tune it on examples of the tasks you need. Put facts that change into retrieval. Training from scratch, as BloombergGPT did with hundreds of billions of tokens, is the most expensive route by far.
Are domain-specific LLMs better than general models?
On their field's tasks they often are when they are built, and the lead can shrink as general models improve on public knowledge. The lasting advantage comes from private knowledge and repeated jobs that no general model was trained on, and from running the model where your data has to stay.
How do you evaluate a domain-specific LLM?
With a held-out set drawn from the field's real work, a protocol fixed before scoring, and confidence intervals on every comparison. Our guide to domain-specific LLM evaluation works through the method using our QMSum study.
What is domain adaptation?
Domain adaptation is the umbrella term for adjusting a general model to one field, through continued pretraining, fine-tuning or both. The glossary entry covers the stages.
The Short Version
A domain-specific LLM knows a field and serves many tasks in it. A task-specific model does one job. The two stack well: domain knowledge first, task skill on top, with retrieval for the facts that change. Public field knowledge keeps flowing into general models, so the lasting value of training sits in your own data and your own repeated work. To go deeper on how knowledge enters a model, read continued pretraining vs fine-tuning vs RAG.
Sources, read 29 October 2026: IBM, "What is a domain-specific LLM?"; Gururangan et al., "Don't Stop Pretraining" (2020); Wu et al., "BloombergGPT" (2023); Colombo et al., "SaulLM-7B" (2024); Labrak et al., "BioMistral" (2024); Cheng et al., "Adapting Large Language Models to Domains via Reading Comprehension" (2023); Cadence, Maestro announcement (26 October 2026); Manufacturers' Monthly, "Choosing between domain-specific and general purpose AI models" (20 October 2026). Our research: the QMSum study.