LMQL

What is LMQL?

LMQL is a programming language for talking to large language models, created by the SRI Lab at ETH Zurich and contributors. A query is written like a Python function: top-level strings are prompts, square-bracket variables such as [ANSWER] mark where the model generates text, and a where clause sets hard constraints that the runtime enforces, for example a length limit, a stop phrase, a set of allowed values, a regex or a type like int. Because queries use ordinary loops, string interpolation and nested queries, prompts become modular programs, and the results come back as plain Python variables. Typed output and dataclasses give schema-safe JSON, and Python functions can be exposed to the model as tools. LMQL runs across backends including OpenAI models, local Hugging Face Transformers models and llama.cpp. It installs with pip, has a browser Playground IDE and a CLI, and is open source with no listed pricing.

Use cases

• Constrain LLM output with a where clause
• Generate schema-safe JSON from a model
• Write multi-step prompts with Python control flow
• Expose Python functions as model tools
• Extract typed values from model responses
• Build reusable prompts with nested queries
• Prototype prompts in the Playground IDE

Who it's for

• Python developers
• LLM researchers
• Prompt engineers
• AI engineers

Why use LMQL

• Hard constraints keep output within a length, set of values, regex or type, instead of relying on the model to follow instructions.
• Typed variables return results as normal Python objects, so downstream code needs less parsing.
• The same query can run on different backends by changing one line.
• Nested queries let you split complex prompts into reusable pieces.
• The Playground shows the decoding graph, which helps you see how a query was executed.

LMQL features

• Constraints on generation with a where clause
• Typed variables such as int and dataclasses for structured output
• Regex constraints on generated text
• Python loops and string interpolation inside prompts
• Nested queries for modular, reusable prompt pieces
• Tool use by exposing Python functions to the model
• Switch between model backends with one line
• Browser-based Playground IDE with a decoding graph view
• Command-line tool, lmql run, for .lmql files
• Query functions callable from Python with the lmql.query decorator

LMQL integrations

• OpenAI
• Azure OpenAI
• llama.cpp
• LangChain
• Pandas

How to get started with LMQL

1. Install LMQL with pip in a Python 3.10 or newer environment, adding the hf extra for local GPU models.
2. Set OPENAI_API_KEY or an api.env file if you plan to use OpenAI models.
3. Write a query: prompt strings, [VARIABLES] for generation and a where clause for constraints.
4. Run it in the Playground with lmql playground, from a .lmql file with lmql run, or as a Python function.
5. Read the generated variables as Python values and chain further queries or tools.

LMQL limitations

• Constraint support is partial on OpenAI Chat API models, which limits what the runtime can enforce
• You need to learn the query syntax
• Local Transformers models need an inference server started with lmql serve-model
• The Playground needs Node.js installed

LMQL pricing

LMQL is open source, so you can use and self-host it for free.

Prices can change and may vary by country or exclude tax. Check the official pricing page.

LMQL alternatives

The main LMQL alternatives listed on Debuly are Guidance, Outlines, Instructor, Pydantic AI, LangChain and BAML. The table compares pricing and platforms; order is A–Z.

1. 

   BAML

   Open-source, TypeScript-like programming language for agents, with typed LLM functions, built-in tests and SDK generation.

   Developer platformsCoding
   Open sourceVisit BAML (official site)
2. 

   Guidance

   Python library for steering language models, combining constrained generation with ordinary control flow so outputs follow regexes, grammars or JSON schemas.

   Developer platforms
   Open sourceVisit Guidance (official site)
3. 

   Instructor

   Open-source library that gets validated, typed structured data out of LLMs using Pydantic models and automatic retries.

   Developer platformsCoding
   Open sourceVisit Instructor (official site)
4. 

   LangChain

   Open-source framework for building LLM agents, paired with LangSmith for tracing, evals and agent deployment.

   Developer platformsAutomation and agentsCoding
   FreemiumVisit LangChain (official site)
5. 

   Outlines

   Python library from .txt that makes LLM output match a type you define, such as a Pydantic model, enum or regex, during generation.

   Developer platformsCodingAutomation and agents
   Open sourceVisit Outlines (official site)
6. 

   Pydantic AI

   A type-safe, model-agnostic Python framework for building agents and LLM applications, from the team behind Pydantic.

   Developer platformsAutomation and agentsCoding
   Open sourceVisit Pydantic AI (official site)

LMQL FAQ

Sources & verification

Facts on this page come from the official sources below, each with the date we read it. Written by Debuly; the vendor did not review this page.

Image credits

• Logo: none collected yet, so a monogram is shown.
• Cover: none collected yet, so a generated tile is shown.

Spot an error? Report an issue with this listing. We also remove a logo or image on the owner's request.