Changelog#

0.30 (2026-03-31)#

  • The register_models() plugin hook now takes an optional model_aliases parameter listing all of the models, async models and aliases that have been registered so far by other plugins. A plugin with @hookimpl(trylast=True) can use this to take previously registered models into account.

  • Added docstrings to public classes and methods and included those directly in the documentation.

0.29 (2026-03-17)#

  • The -t/--template option now works correctly with the -x/--extract and --xl/--extract-last flags.

  • llm logs now shows any additional model options in the Markdown output. #1322

  • New OpenAI models: gpt-5.4, gpt-5.4-mini, gpt-5.4-nano. #1376

0.28 (2025-12-12)#

  • New OpenAI models: gpt-5.1, gpt-5.1-chat-latest, gpt-5.2 and gpt-5.2-chat-latest. #1300, #1317

  • LLM now requires Python 3.10 or higher. Python 3.14 is now covered by the tests.

  • When fetching URLs as fragments using llm -f URL, the request now includes a custom user-agent header: llm/VERSION (https://llm.datasette.io/). #1309

  • Fixed a bug where fragments were not correctly registered with their source when using llm chat. Thanks, Giuseppe Rota. #1316

  • Fixed some file descriptor leak warnings. Thanks, Eric Bloch. #1313

  • Fixed a deprecation warning for asyncio.iscoroutinefunction.

  • Type annotations for the OpenAI Chat, AsyncChat and Completion execute() methods. Thanks, Arjan Mossel. #1315

  • The project now uses uv and dependency groups for development. See the updated contributing documentation. #1318

0.27.1 (2025-08-11)#

  • llm chat -t template now correctly loads any tools that are included in that template. #1239

  • Fixed a bug where llm -m gpt5 -o reasoning_effort minimal --save gm saved a template containing invalid YAML. #1237

  • Fixed a bug where running llm chat -t template could cause prompts to be duplicated. #1240

  • Less confusing error message if a requested toolbox class is unavailable. #1238

0.27 (2025-08-11)#

This release adds support for the new GPT-5 family of models from OpenAI. It also enhances tool calling in a number of ways, including allowing templates to bundle pre-configured tools.

New features#

  • New models: gpt-5, gpt-5-mini and gpt-5-nano. #1229

  • LLM templates can now include a list of tools. These can be named tools from plugins or arbitrary Python function blocks, see Tools in templates. #1009

  • Tools can now return attachments, for models that support features such as image input. #1014

  • New methods on the Toolbox class: .add_tool(), .prepare() and .prepare_async(), described in Dynamic toolboxes. #1111

  • New model.conversation(before_call=x, after_call=y) attributes for registering callback functions to run before and after tool calls. See tool debugging hooks for details. #1088

  • Some model providers can serve different models from the same configured URL - llm-llama-server for example. Plugins for these providers can now record the resolved model ID of the model that was used to the LLM logs using the response.set_resolved_model(model_id) method. #1117

  • Raising llm.CancelToolCall now only cancels the current tool call, passing an error back to the model and allowing it to continue. #1148

  • New -l/--latest option for llm logs -q searchterm for searching logs ordered by date (most recent first) instead of the default relevance search. #1177

Bug fixes and documentation#

  • Fix for various bugs with different formats of streaming function responses for OpenAI-compatible models. Thanks, James Sanford. #1218

  • The register_embedding_models hook is now documented. #1049

  • Show visible stack trace for llm templates show invalid-template-name. #1053

  • Handle invalid tool names more gracefully in llm chat. #1104

  • Add a Tool plugins section to the plugin directory. #1110

  • Error on register(Klass) if the passed class is not a subclass of Toolbox. #1114

  • Add -h for --help for all llm CLI commands. #1134

  • Add missing dataclasses to advanced model plugins docs. #1137

  • Fixed a bug where llm logs -T llm_version "version" --async incorrectly recorded just one single log entry when it should have recorded two. #1150

  • All extra OpenAI model keys in extra-openai-models.yaml are now documented. #1228

0.26 (2025-05-27)#

Tool support is finally here! This release adds support exposing tools to LLMs, previously described in the release notes for 0.26a0 and 0.26a1.

Read Large Language Models can run tools in your terminal with LLM 0.26 for a detailed overview of the new features.

Also in this release:

0.26a1 (2025-05-25)#

Hopefully the last alpha before a stable release that includes tool support.

Features#

  • Plugin-provided tools can now be grouped into “Toolboxes”.

    • Toolboxes (llm.Toolbox classes) allow plugins to expose multiple related tools that share state or configuration (e.g., a Memory tool or Filesystem tool). (#1059, #1086)

  • Tool support for llm chat.

    • The llm chat command now accepts --tool and --functions arguments, allowing interactive chat sessions to use tools. (#1004, #1062)

  • Tools can now execute asynchronously.

    • Models that implement AsyncModel can now run tools, including tool functions defined as async def. (#1063)

  • llm chat now supports adding fragments during a session.

    • Use the new !fragment <id> command while chatting to insert content from a fragment. Initial fragments can also be passed to llm chat using -f or --sf. Thanks, Dan Turkel. (#1044, #1048)

  • Filter llm logs by tools.

    • New --tool <name> option to filter logs to show only responses that involved a specific tool (e.g., --tool simple_eval).

    • The --tools flag shows all responses that used any tool. (#1013, #1072)

  • llm schemas list can output JSON.

    • Added --json and --nl (newline-delimited JSON) options to llm schemas list for programmatic access to saved schema definitions. (#1070)

  • Filter llm similar results by ID prefix.

    • The new --prefix option for llm similar allows searching for similar items only within IDs that start with a specified string (e.g., llm similar my-collection --prefix 'docs/'). Thanks, Dan Turkel. (#1052)

  • Control chained tool execution limit.

    • New --chain-limit <N> (or --cl) option for llm prompt and llm chat to specify the maximum number of consecutive tool calls allowed for a single prompt. Defaults to 5; set to 0 for unlimited. (#1025)

  • llm plugins --hook <NAME> option.

    • Filter the list of installed plugins to only show those that implement a specific plugin hook. (#1047)

  • llm tools list now shows toolboxes and their methods. (#1013)

  • llm prompt and llm chat now automatically re-enable plugin-provided tools when continuing a conversation (-c or --cid). (#1020)

  • The --tools-debug option now pretty-prints JSON tool results for improved readability. (#1083)

  • New LLM_TOOLS_DEBUG environment variable to permanently enable --tools-debug. (#1045)

  • llm chat sessions now correctly respect default model options configured with llm models set-options. Thanks, André Arko. (#985)

  • New --pre option for llm install to allow installing pre-release packages. (#1060)

  • OpenAI models (gpt-4o, gpt-4o-mini) now explicitly declare support for tools and vision. (#1037)

  • The supports_tools parameter is now supported in extra-openai-models.yaml. Thanks, Mahesh Hegde . (#1068)

Bug fixes#

  • Fixed a bug where the name parameter in register(function, name="name") was ignored for tool plugins. (#1032)

  • Ensure pathlib.Path objects are cast to str before passing to click.edit in llm templates edit. Thanks, Abizer Lokhandwala. (#1031)

0.26a0 (2025-05-13)#

This is the first alpha to introduce support for tools! Models with tool capability (which includes the default OpenAI model family) can now be granted access to execute Python functions as part of responding to a prompt.

Tools are supported by the command-line interface:

llm --functions '
def multiply(x: int, y: int) -> int:
    """Multiply two numbers."""
    return x * y
' 'what is 34234 * 213345'

And in the Python API, using a new model.chain() method for executing multiple prompts in a sequence:

import llm

def multiply(x: int, y: int) -> int:
    """Multiply two numbers."""
    return x * y

model = llm.get_model("gpt-4.1-mini")
response = model.chain(
    "What is 34234 * 213345?",
    tools=[multiply]
)
print(response.text())

New tools can also be defined using the register_tools() plugin hook. They can then be called by name from the command-line like this:

llm -T multiply 'What is 34234 * 213345?'

Tool support is currently under active development. Consult this milestone for the latest status.

0.25 (2025-05-04)#

  • New plugin feature: register_fragment_loaders(register) plugins can now return a mixture of fragments and attachments. The llm-video-frames plugin is the first to take advantage of this mechanism. #972

  • New OpenAI models: gpt-4.1, gpt-4.1-mini, gpt-41-nano, o3, o4-mini. #945, #965, #976.

  • New environment variables: LLM_MODEL and LLM_EMBEDDING_MODEL for setting the model to use without needing to specify -m model_id every time. #932

  • New command: llm fragments loaders, to list all currently available fragment loader prefixes provided by plugins. #941

  • llm fragments command now shows fragments ordered by the date they were first used. #973

  • llm chat now includes a !edit command for editing a prompt using your default terminal text editor. Thanks, Benedikt Willi. #969

  • Allow -t and --system to be used at the same time. #916

  • Fixed a bug where accessing a model via its alias would fail to respect any default options set for that model. #968

  • Improved documentation for extra-openai-models.yaml. Thanks, Rahim Nathwani and Dan Guido. #950, #957

  • llm -c/--continue now works correctly with the -d/--database option. llm chat now accepts that -d/--database option. Thanks, Sukhbinder Singh. #933

0.25a0 (2025-04-10)#

  • llm models --options now shows keys and environment variables for models that use API keys. Thanks, Steve Morin. #903

  • Added py.typed marker file so LLM can now be used as a dependency in projects that use mypy without a warning. #887

  • $ characters can now be used in templates by escaping them as $$. Thanks, @guspix. #904

  • LLM now uses pyproject.toml instead of setup.py. #908

0.24.2 (2025-04-08)#

  • Fixed a bug on Windows with the new llm -t path/to/file.yaml feature. #901

0.24.1 (2025-04-08)#

  • Templates can now be specified as a path to a file on disk, using llm -t path/to/file.yaml. This makes them consistent with how -f fragments are loaded. #897

  • llm logs backup /tmp/backup.db command for backing up your logs.db database. #879

0.24 (2025-04-07)#

Support for fragments to help assemble prompts for long context models. Improved support for templates to support attachments and fragments. New plugin hooks for providing custom loaders for both templates and fragments. See Long context support in LLM 0.24 using fragments and template plugins for more on this release.

The new llm-docs plugin demonstrates these new features. Install it like this:

llm install llm-docs

Now you can ask questions of the LLM documentation like this:

llm -f docs: 'How do I save a new template?'

The docs: prefix is registered by the plugin. The plugin fetches the LLM documentation for your installed version (from the docs-for-llms repository) and uses that as a prompt fragment to help answer your question.

Two more new plugins are llm-templates-github and llm-templates-fabric.

llm-templates-github lets you share and use templates on GitHub. You can run my Pelican riding a bicycle benchmark against a model like this:

llm install llm-templates-github
llm -t gh:simonw/pelican-svg -m o3-mini

This executes this pelican-svg.yaml template stored in my simonw/llm-templates repository, using a new repository naming convention.

To share your own templates, create a repository on GitHub under your user account called llm-templates and start saving .yaml files to it.

llm-templates-fabric provides a similar mechanism for loading templates from Daniel Miessler’s fabric collection:

llm install llm-templates-fabric
curl https://simonwillison.net/2025/Apr/6/only-miffy/ | \
  llm -t f:extract_main_idea

Major new features:

  • New fragments feature. Fragments can be used to assemble long prompts from multiple existing pieces - URLs, file paths or previously used fragments. These will be stored de-duplicated in the database avoiding wasting space storing multiple long context pieces. Example usage: llm -f https://llm.datasette.io/robots.txt 'explain this file'. #617

  • The llm logs file now accepts -f fragment references too, and will show just logged prompts that used those fragments.

  • register_template_loaders() plugin hook allowing plugins to register new prefix:value custom template loaders. #809

  • register_fragment_loaders() plugin hook allowing plugins to register new prefix:value custom fragment loaders. #886

  • llm fragments family of commands for browsing fragments that have been previously logged to the database.

  • The new llm-openai plugin provides support for o1-pro (which is not supported by the OpenAI mechanism used by LLM core). Future OpenAI features will migrate to this plugin instead of LLM core itself.

Improvements to templates:

  • llm -t $URL option can now take a URL to a YAML template. #856

  • Templates can now store default model options. #845

  • Executing a template that does not use the $input variable no longer blocks LLM waiting for input, so prompt templates can now be used to try different models using llm -t pelican-svg -m model_id. #835

  • llm templates command no longer crashes if one of the listed template files contains invalid YAML. #880

  • Attachments can now be stored in templates. #826

Other changes:

  • New llm models options family of commands for setting default options for particular models. #829

  • llm logs list, llm schemas list and llm schemas show all now take a -d/--database option with an optional path to a SQLite database. They used to take -p/--path but that was inconsistent with other commands. -p/--path still works but is excluded from --help and will be removed in a future LLM release. #857

  • llm logs -e/--expand option for expanding fragments. #881

  • llm prompt -d path-to-sqlite.db option can now be used to write logs to a custom SQLite database. #858

  • llm similar -p/--plain option providing more human-readable output than the default JSON. #853

  • llm logs -s/--short now truncates to include the end of the prompt too. Thanks, Sukhbinder Singh. #759

  • Set the LLM_RAISE_ERRORS=1 environment variable to raise errors during prompts rather than suppressing them, which means you can run python -i -m llm 'prompt' and then drop into a debugger on errors with import pdb; pdb.pm(). #817

  • Improved –help output for llm embed-multi. #824

  • llm models -m X option which can be passed multiple times with model IDs to see the details of just those models. #825

  • OpenAI models now accept PDF attachments. #834

  • llm prompt -q gpt -q 4o option - pass -q searchterm one or more times to execute a prompt against the first model that matches all of those strings - useful for if you can’t remember the full model ID. #841

  • OpenAI compatible models configured using extra-openai-models.yaml now support supports_schema: true, vision: true and audio: true options. Thanks @adaitche and @giuli007. #819, #843

0.24a1 (2025-04-06)#

  • New Fragments feature. #617

  • register_fragment_loaders() plugin hook. #809

0.24a0 (2025-02-28)#

  • Alpha release with experimental register_template_loaders() plugin hook. #809

0.23 (2025-02-28)#

Support for schemas, for getting supported models to output JSON that matches a specified JSON schema. See also Structured data extraction from unstructured content using LLM schemas for background on this feature. #776

  • New llm prompt --schema '{JSON schema goes here} option for specifying a schema that should be used for the output from the model. The schemas documentation has more details and a tutorial.

  • Schemas can also be defined using a concise schema specification, for example llm prompt --schema 'name, bio, age int'. #790

  • Schemas can also be specified by passing a filename and through several other methods. #780

  • New llm schemas family of commands: llm schemas list, llm schemas show, and llm schemas dsl for debugging the new concise schema language. #781

  • Schemas can now be saved to templates using llm --schema X --save template-name or through modifying the template YAML. #778

  • The llm logs command now has new options for extracting data collected using schemas: --data, --data-key, --data-array, --data-ids. #782

  • New llm logs --id-gt X and --id-gte X options. #801

  • New llm models --schemas option for listing models that support schemas. #797

  • model.prompt(..., schema={...}) parameter for specifying a schema from Python. This accepts either a dictionary JSON schema definition or a Pydantic BaseModel subclass, see schemas in the Python API docs.

  • The default OpenAI plugin now enables schemas across all supported models. Run llm models --schemas for a list of these.

  • The llm-anthropic and llm-gemini plugins have been upgraded to add schema support for those models. Here’s documentation on how to add schema support to a model plugin.

Other smaller changes:

  • GPT-4.5 preview is now a supported model: llm -m gpt-4.5 'a joke about a pelican and a wolf' #795

  • The prompt string is now optional when calling model.prompt() from the Python API, so model.prompt(attachments=llm.Attachment(url=url))) now works. #784

  • extra-openai-models.yaml now supports a reasoning: true option. Thanks, Kasper Primdal Lauritzen. #766

  • LLM now depends on Pydantic v2 or higher. Pydantic v1 is no longer supported. #520

0.22 (2025-02-16)#

See also LLM 0.22, the annotated release notes.

  • Plugins that provide models that use API keys can now subclass the new llm.KeyModel and llm.AsyncKeyModel classes. This results in the API key being passed as a new key parameter to their .execute() methods, and means that Python users can pass a key as the model.prompt(..., key=) - see Passing an API key. Plugin developers should consult the new documentation on writing Models that accept API keys. #744

  • New OpenAI model: chatgpt-4o-latest. This model ID accesses the current model being used to power ChatGPT, which can change without warning. #752

  • New llm logs -s/--short flag, which returns a greatly shortened version of the matching log entries in YAML format with a truncated prompt and without including the response. #737

  • Both llm models and llm embed-models now take multiple -q search fragments. You can now search for all models matching “gemini” and “exp” using llm models -q gemini -q exp. #748

  • New llm embed-multi --prepend X option for prepending a string to each value before it is embedded - useful for models such as nomic-embed-text-v2-moe that require passages to start with a string like "search_document: ". #745

  • The response.json() and response.usage() methods are now documented.

  • Fixed a bug where conversations that were loaded from the database could not be continued using asyncio prompts. #742

  • New plugin for macOS users: llm-mlx, which provides extremely high performance access to a wide range of local models using Apple’s MLX framework.

  • The llm-claude-3 plugin has been renamed to llm-anthropic.

0.21 (2025-01-31)#

  • New model: o3-mini. #728

  • The o3-mini and o1 models now support a reasoning_effort option which can be set to low, medium or high.

  • llm prompt and llm logs now have a --xl/--extract-last option for extracting the last fenced code block in the response - a complement to the existing --x/--extract option. #717

0.20 (2025-01-22)#

  • New model, o1. This model does not yet support streaming. #676

  • o1-preview and o1-mini models now support streaming.

  • New models, gpt-4o-audio-preview and gpt-4o-mini-audio-preview. #677

  • llm prompt -x/--extract option, which returns just the content of the first fenced code block in the response. Try llm prompt -x 'Python function to reverse a string'. #681

    • Creating a template using llm ... --save x now supports the -x/--extract option, which is saved to the template. YAML templates can set this option using extract: true.

    • New llm logs -x/--extract option extracts the first fenced code block from matching logged responses.

  • New llm models -q 'search' option returning models that case-insensitively match the search query. #700

  • Installation documentation now also includes uv. Thanks, Ariel Marcus. #690 and #702

  • llm models command now shows the current default model at the bottom of the listing. Thanks, Amjith Ramanujam. #688

  • Plugin directory now includes llm-venice, llm-bedrock, llm-deepseek and llm-cmd-comp.

  • Fixed bug where some dependency version combinations could cause a Client.__init__() got an unexpected keyword argument 'proxies' error. #709

  • OpenAI embedding models are now available using their full names of text-embedding-ada-002, text-embedding-3-small and text-embedding-3-large - the previous names are still supported as aliases. Thanks, web-sst. #654

0.19.1 (2024-12-05)#

  • FIxed bug where llm.get_models() and llm.get_async_models() returned the same model multiple times. #667

0.19 (2024-12-01)#

  • Tokens used by a response are now logged to new input_tokens and output_tokens integer columns and a token_details JSON string column, for the default OpenAI models and models from other plugins that implement this feature. #610

  • llm prompt now takes a -u/--usage flag to display token usage at the end of the response.

  • llm logs -u/--usage shows token usage information for logged responses.

  • llm prompt ... --async responses are now logged to the database. #641

  • llm.get_models() and llm.get_async_models() functions, documented here. #640

  • response.usage() and async response await response.usage() methods, returning a Usage(input=2, output=1, details=None) dataclass. #644

  • response.on_done(callback) and await response.on_done(callback) methods for specifying a callback to be executed when a response has completed, documented here. #653

  • Fix for bug running llm chat on Windows 11. Thanks, Sukhbinder Singh. #495

0.19a2 (2024-11-20)#

0.19a1 (2024-11-19)#

  • response.usage() and async response await response.usage() methods, returning a Usage(input=2, output=1, details=None) dataclass. #644

0.19a0 (2024-11-19)#

  • Tokens used by a response are now logged to new input_tokens and output_tokens integer columns and a token_details JSON string column, for the default OpenAI models and models from other plugins that implement this feature. #610

  • llm prompt now takes a -u/--usage flag to display token usage at the end of the response.

  • llm logs -u/--usage shows token usage information for logged responses.

  • llm prompt ... --async responses are now logged to the database. #641

0.18 (2024-11-17)#

  • Initial support for async models. Plugins can now provide an AsyncModel subclass that can be accessed in the Python API using the new llm.get_async_model(model_id) method. See async models in the Python API docs and implementing async models in plugins. #507

  • OpenAI models all now include async models, so function calls such as llm.get_async_model("gpt-4o-mini") will return an async model.

  • gpt-4o-audio-preview model can be used to send audio attachments to the GPT-4o audio model. #608

  • Attachments can now be sent without requiring a prompt. #611

  • llm models --options now includes information on whether a model supports attachments. #612

  • llm models --async shows available async models.

  • Custom OpenAI-compatible models can now be marked as can_stream: false in the YAML if they do not support streaming. Thanks, Chris Mungall. #600

  • Fixed bug where OpenAI usage data was incorrectly serialized to JSON. #614

  • Standardized on audio/wav MIME type for audio attachments rather than audio/wave. #603

0.18a1 (2024-11-14)#

  • Fixed bug where conversations did not work for async OpenAI models. #632

  • __repr__ methods for Response and AsyncResponse.

0.18a0 (2024-11-13)#

Alpha support for async models. #507

Multiple smaller changes.

0.17 (2024-10-29)#

Support for attachments, allowing multi-modal models to accept images, audio, video and other formats. #578

The default OpenAI gpt-4o and gpt-4o-mini models can both now be prompted with JPEG, GIF, PNG and WEBP images.

Attachments in the CLI can be URLs:

llm -m gpt-4o "describe this image" \
  -a https://static.simonwillison.net/static/2024/pelicans.jpg

Or file paths:

llm -m gpt-4o-mini "extract text" -a image1.jpg -a image2.jpg

Or binary data, which may need to use --attachment-type to specify the MIME type:

cat image | llm -m gpt-4o-mini "extract text" --attachment-type - image/jpeg

Attachments are also available in the Python API:

model = llm.get_model("gpt-4o-mini")
response = model.prompt(
    "Describe these images",
    attachments=[
        llm.Attachment(path="pelican.jpg"),
        llm.Attachment(url="https://static.simonwillison.net/static/2024/pelicans.jpg"),
    ]
)

Plugins that provide alternative models can support attachments, see Attachments for multi-modal models for details.

The latest llm-claude-3 plugin now supports attachments for Anthropic’s Claude 3 and 3.5 models. The llm-gemini plugin supports attachments for Google’s Gemini 1.5 models.

Also in this release: OpenAI models now record their "usage" data in the database even when the response was streamed. These records can be viewed using llm logs --json. #591

0.17a0 (2024-10-28)#

Alpha support for attachments. #578

0.16 (2024-09-12)#

  • OpenAI models now use the internal self.get_key() mechanism, which means they can be used from Python code in a way that will pick up keys that have been configured using llm keys set or the OPENAI_API_KEY environment variable. #552. This code now works correctly:

    import llm
print(llm.get_model("gpt-4o-mini").prompt("hi"))

  • New documented API methods: llm.get_default_model(), llm.set_default_model(alias), llm.get_default_embedding_model(alias), llm.set_default_embedding_model(). #553

  • Support for OpenAI’s new o1 family of preview models, llm -m o1-preview "prompt" and llm -m o1-mini "prompt". These models are currently only available to tier 5 OpenAI API users, though this may change in the future. #570

0.15 (2024-07-18)#

  • Support for OpenAI’s new GPT-4o mini model: llm -m gpt-4o-mini 'rave about pelicans in French' #536

  • gpt-4o-mini is now the default model if you do not specify your own default, replacing GPT-3.5 Turbo. GPT-4o mini is both cheaper and better than GPT-3.5 Turbo.

  • Fixed a bug where llm logs -q 'flourish' -m haiku could not combine both the -q search query and the -m model specifier. #515

0.14 (2024-05-13)#

  • Support for OpenAI’s new GPT-4o model: llm -m gpt-4o 'say hi in Spanish' #490

  • The gpt-4-turbo alias is now a model ID, which indicates the latest version of OpenAI’s GPT-4 Turbo text and image model. Your existing logs.db database may contain records under the previous model ID of gpt-4-turbo-preview. #493

  • New llm logs -r/--response option for outputting just the last captured response, without wrapping it in Markdown and accompanying it with the prompt. #431

  • Nine new plugins since version 0.13:

0.13.1 (2024-01-26)#

  • Fix for No module named 'readline' error on Windows. #407

0.13 (2024-01-26)#

See also LLM 0.13: The annotated release notes.

  • Added support for new OpenAI embedding models: 3-small and 3-large and three variants of those with different dimension sizes, 3-small-512, 3-large-256 and 3-large-1024. See OpenAI embedding models for details. #394

  • The default gpt-4-turbo model alias now points to gpt-4-turbo-preview, which uses the most recent OpenAI GPT-4 turbo model (currently gpt-4-0125-preview). #396

  • New OpenAI model aliases gpt-4-1106-preview and gpt-4-0125-preview.

  • OpenAI models now support a -o json_object 1 option which will cause their output to be returned as a valid JSON object. #373

  • New plugins since the last release include llm-mistral, llm-gemini, llm-ollama and llm-bedrock-meta.

  • The keys.json file for storing API keys is now created with 600 file permissions. #351

  • Documented a pattern for installing plugins that depend on PyTorch using the Homebrew version of LLM, despite Homebrew using Python 3.12 when PyTorch have not yet released a stable package for that Python version. #397

  • Underlying OpenAI Python library has been upgraded to >1.0. It is possible this could cause compatibility issues with LLM plugins that also depend on that library. #325

  • Arrow keys now work inside the llm chat command. #376

  • LLM_OPENAI_SHOW_RESPONSES=1 environment variable now outputs much more detailed information about the HTTP request and response made to OpenAI (and OpenAI-compatible) APIs. #404

  • Dropped support for Python 3.7.

0.12 (2023-11-06)#

  • Support for the new GPT-4 Turbo model from OpenAI. Try it using llm chat -m gpt-4-turbo or llm chat -m 4t. #323

  • New -o seed 1 option for OpenAI models which sets a seed that can attempt to evaluate the prompt deterministically. #324

0.11.2 (2023-11-06)#

  • Pin to version of OpenAI Python library prior to 1.0 to avoid breaking. #327

0.11.1 (2023-10-31)#

0.11 (2023-09-18)#

LLM now supports the new OpenAI gpt-3.5-turbo-instruct model, and OpenAI completion (as opposed to chat completion) models in general. #284

llm -m gpt-3.5-turbo-instruct 'Reasons to tame a wild beaver:'

OpenAI completion models like this support a -o logprobs 3 option, which accepts a number between 1 and 5 and will include the log probabilities (for each produced token, what were the top 3 options considered by the model) in the logged response.

llm -m gpt-3.5-turbo-instruct 'Say hello succinctly' -o logprobs 3

You can then view the logprobs that were recorded in the SQLite logs database like this:

sqlite-utils "$(llm logs path)" \
  'select * from responses order by id desc limit 1' | \
  jq '.[0].response_json' -r | jq

Truncated output looks like this:

  [
    {
      "text": "Hi",
      "top_logprobs": [
        {
          "Hi": -0.13706253,
          "Hello": -2.3714375,
          "Hey": -3.3714373
        }
      ]
    },
    {
      "text": " there",
      "top_logprobs": [
        {
          " there": -0.96057636,
          "!\"": -0.5855763,
          ".\"": -3.2574513
        }
      ]
    }
  ]

Also in this release:

  • The llm.user_dir() function, used by plugins, now ensures the directory exists before returning it. #275

  • New LLM_OPENAI_SHOW_RESPONSES=1 environment variable for displaying the full HTTP response returned by OpenAI compatible APIs. #286

  • The llm embed-multi command now has a --batch-size X option for setting the batch size to use when processing embeddings - useful if you have limited memory available. #273

  • The collection.embed_multi() method also now accepts an optional batch_size=int argument.

  • Fixed two bugs with llm embed-multi --files relating to handling of directories. Thanks, ealvar3z. #274, #280

0.10 (2023-09-12)#

The two major features in this release are the llm chat command and support for embedding binary data.

See Build an image search engine with llm-clip, chat with models with llm chat for more background on these features.

llm chat#

The new llm chat command starts an ongoing chat conversation with a model in your terminal. It works with all models supported by LLM and its plugins, including locally installed models such as Llama 2. #231

This offers a big performance boost for local models, since they don’t need to be freshly loaded into memory for each prompt.

Here’s an example chat with Llama 2 13B, provided by the llm-mlc plugin.

llm chat -m mlc-chat-Llama-2-13b-chat-hf-q4f16_1

This starts a chat session:

Type 'exit' or 'quit' to exit
Type '!multi' to enter multiple lines, then '!end' to finish
Type '!edit' to open your default editor and modify the prompt.
> Who are you?
Hello! I'm just an AI, here to assist you with any questions you may have.
My name is LLaMA, and I'm a large language model trained to provide helpful
and informative responses to a wide range of questions and topics. I'm here
to help you find the information you need, while ensuring a safe and
respectful conversation. Please feel free to ask me anything, and I'll do my
best to provide a helpful and accurate response.
> Tell me a joke about otters
Sure, here's a joke about otters:

Why did the otter go to the party?

Because he heard it was a "whale" of a time!

(Get it? Whale, like a big sea mammal, but also a "wild" or "fun" time.
Otters are known for their playful and social nature, so it's a lighthearted
and silly joke.)

I hope that brought a smile to your face! Do you have any other questions or
topics you'd like to discuss?
> exit

Chat sessions are logged to SQLite - use llm logs to view them. They can accept system prompts, templates and model options - consult the chat documentation for details.

Binary embedding support#

LLM’s embeddings feature has been expanded to provide support for embedding binary data, in addition to text. #254

This enables models like CLIP, supported by the new llm-clip plugin.

CLIP is a multi-modal embedding model which can embed images and text into the same vector space. This means you can use it to create an embedding index of photos, and then search for the embedding vector for “a happy dog” and get back images that are semantically closest to that string.

To create embeddings for every JPEG in a directory stored in a photos collection, run:

llm install llm-clip
llm embed-multi photos --files photos/ '*.jpg' --binary -m clip

Now you can search for photos of raccoons using:

llm similar photos -c 'raccoon'

This spits out a list of images, ranked by how similar they are to the string “raccoon”:

{"id": "IMG_4801.jpeg", "score": 0.28125139257127457, "content": null, "metadata": null}
{"id": "IMG_4656.jpeg", "score": 0.26626441704164294, "content": null, "metadata": null}
{"id": "IMG_2944.jpeg", "score": 0.2647445926996852, "content": null, "metadata": null}
...

Also in this release#

  • The LLM_LOAD_PLUGINS environment variable can be used to control which plugins are loaded when llm starts running. #256

  • The llm plugins --all option includes builtin plugins in the list of plugins. #259

  • The llm embed-db family of commands has been renamed to llm collections. #229

  • llm embed-multi --files now has an --encoding option and defaults to falling back to latin-1 if a file cannot be processed as utf-8. #225

0.10a1 (2023-09-11)#

  • Support for embedding binary data. #254

  • llm chat now works for models with API keys. #247

  • llm chat -o for passing options to a model. #244

  • llm chat --no-stream option. #248

  • LLM_LOAD_PLUGINS environment variable. #256

  • llm plugins --all option for including builtin plugins. #259

  • llm embed-db has been renamed to llm collections. #229

  • Fixed bug where llm embed -c option was treated as a filepath, not a string. Thanks, mhalle. #263

0.10a0 (2023-09-04)#

  • New llm chat command for starting an interactive terminal chat with a model. #231

  • llm embed-multi --files now has an --encoding option and defaults to falling back to latin-1 if a file cannot be processed as utf-8. #225

0.9 (2023-09-03)#

The big new feature in this release is support for embeddings. See LLM now provides tools for working with embeddings for additional details.

Embedding models take a piece of text - a word, sentence, paragraph or even a whole article, and convert that into an array of floating point numbers. #185

This embedding vector can be thought of as representing a position in many-dimensional-space, where the distance between two vectors represents how semantically similar they are to each other within the content of a language model.

Embeddings can be used to find related documents, and also to implement semantic search - where a user can search for a phrase and get back results that are semantically similar to that phrase even if they do not share any exact keywords.

LLM now provides both CLI and Python APIs for working with embeddings. Embedding models are defined by plugins, so you can install additional models using the plugins mechanism.

The first two embedding models supported by LLM are:

See Embedding with the CLI for detailed instructions on working with embeddings using LLM.

The new commands for working with embeddings are:

  • llm embed - calculate embeddings for content and return them to the console or store them in a SQLite database.

  • llm embed-multi - run bulk embeddings for multiple strings, using input from a CSV, TSV or JSON file, data from a SQLite database or data found by scanning the filesystem. #215

  • llm similar - run similarity searches against your stored embeddings - starting with a search phrase or finding content related to a previously stored vector. #190

  • llm embed-models - list available embedding models.

  • llm embed-db - commands for inspecting and working with the default embeddings SQLite database.

There’s also a new llm.Collection class for creating and searching collections of embedding from Python code, and a llm.get_embedding_model() interface for embedding strings directly. #191

0.8.1 (2023-08-31)#

  • Fixed bug where first prompt would show an error if the io.datasette.llm directory had not yet been created. #193

  • Updated documentation to recommend a different llm-gpt4all model since the one we were using is no longer available. #195

0.8 (2023-08-20)#

  • The output format for llm logs has changed. Previously it was JSON - it’s now a much more readable Markdown format suitable for pasting into other documents. #160

    • The new llm logs --json option can be used to get the old JSON format.

    • Pass llm logs --conversation ID or --cid ID to see the full logs for a specific conversation.

  • You can now combine piped input and a prompt in a single command: cat script.py | llm 'explain this code'. This works even for models that do not support system prompts. #153

  • Additional OpenAI-compatible models can now be configured with custom HTTP headers. This enables platforms such as openrouter.ai to be used with LLM, which can provide Claude access even without an Anthropic API key.

  • Keys set in keys.json are now used in preference to environment variables. #158

  • The documentation now includes a plugin directory listing all available plugins for LLM. #173

  • New related tools section in the documentation describing ttok, strip-tags and symbex. #111

  • The llm models, llm aliases and llm templates commands now default to running the same command as llm models list and llm aliases list and llm templates list. #167

  • New llm keys (aka llm keys list) command for listing the names of all configured keys. #174

  • Two new Python API functions, llm.set_alias(alias, model_id) and llm.remove_alias(alias) can be used to configure aliases from within Python code. #154

  • LLM is now compatible with both Pydantic 1 and Pydantic 2. This means you can install llm as a Python dependency in a project that depends on Pydantic 1 without running into dependency conflicts. Thanks, Chris Mungall. #147

  • llm.get_model(model_id) is now documented as raising llm.UnknownModelError if the requested model does not exist. #155

0.7.1 (2023-08-19)#

  • Fixed a bug where some users would see an AlterError: No such column: log.id error when attempting to use this tool, after upgrading to the latest sqlite-utils 3.35 release. #162

0.7 (2023-08-12)#

The new Model aliases commands can be used to configure additional aliases for models, for example:

llm aliases set turbo gpt-3.5-turbo-16k

Now you can run the 16,000 token gpt-3.5-turbo-16k model like this:

llm -m turbo 'An epic Greek-style saga about a cheesecake that builds a SQL database from scratch'

Use llm aliases list to see a list of aliases and llm aliases remove turbo to remove one again. #151

Notable new plugins#

Also in this release#

  • OpenAI models now have min and max validation on their floating point options. Thanks, Pavel Král. #115

  • Fix for bug where llm templates list raised an error if a template had an empty prompt. Thanks, Sherwin Daganato. #132

  • Fixed bug in llm install --editable option which prevented installation of .[test]. #136

  • llm install --no-cache-dir and --force-reinstall options. #146

0.6.1 (2023-07-24)#

  • LLM can now be installed directly from Homebrew core: brew install llm. #124

  • Python API documentation now covers System prompts.

  • Fixed incorrect example in the Templates documentation. Thanks, Jorge Cabello. #125

0.6 (2023-07-18)#

  • Models hosted on Replicate can now be accessed using the llm-replicate plugin, including the new Llama 2 model from Meta AI. More details here: Accessing Llama 2 from the command-line with the llm-replicate plugin.

  • Model providers that expose an API that is compatible with the OpenAPI API format, including self-hosted model servers such as LocalAI, can now be accessed using additional configuration for the default OpenAI plugin. #106

  • OpenAI models that are not yet supported by LLM can also be configured using the new extra-openai-models.yaml configuration file. #107

  • The llm logs command now accepts a -m model_id option to filter logs to a specific model. Aliases can be used here in addition to model IDs. #108

  • Logs now have a SQLite full-text search index against their prompts and responses, and the llm logs -q SEARCH option can be used to return logs that match a search term. #109

0.5 (2023-07-12)#

LLM now supports additional language models, thanks to a new plugins mechanism for installing additional models.

Plugins are available for 19 models in addition to the default OpenAI ones:

  • llm-gpt4all adds support for 17 models that can download and run on your own device, including Vicuna, Falcon and wizardLM.

  • llm-mpt30b adds support for the MPT-30B model, a 19GB download.

  • llm-palm adds support for Google’s PaLM 2 via the Google API.

A comprehensive tutorial, writing a plugin to support a new model describes how to add new models by building plugins in detail.

New features#

  • Python API documentation for using LLM models, including models from plugins, directly from Python. #75

  • Messages are now logged to the database by default - no need to run the llm init-db command any more, which has been removed. Instead, you can toggle this behavior off using llm logs off or turn it on again using llm logs on. The llm logs status command shows the current status of the log database. If logging is turned off, passing --log to the llm prompt command will cause that prompt to be logged anyway. #98

  • New database schema for logged messages, with conversations and responses tables. If you have previously used the old logs table it will continue to exist but will no longer be written to. #91

  • New -o/--option name value syntax for setting options for models, such as temperature. Available options differ for different models. #63

  • llm models list --options command for viewing all available model options. #82

  • llm "prompt" --save template option for saving a prompt directly to a template. #55

  • Prompt templates can now specify default values for parameters. Thanks, Chris Mungall. #57

  • llm openai models command to list all available OpenAI models from their API. #70

  • llm models default MODEL_ID to set a different model as the default to be used when llm is run without the -m/--model option. #31

Smaller improvements#

  • llm -s is now a shortcut for llm --system. #69

  • llm -m 4-32k alias for gpt-4-32k.

  • llm install -e directory command for installing a plugin from a local directory.

  • The LLM_USER_PATH environment variable now controls the location of the directory in which LLM stores its data. This replaces the old LLM_KEYS_PATH and LLM_LOG_PATH and LLM_TEMPLATES_PATH variables. #76

  • Documentation covering Utility functions for plugins.

  • Documentation site now uses Plausible for analytics. #79

0.4.1 (2023-06-17)#

  • LLM can now be installed using Homebrew: brew install simonw/llm/llm. #50

  • llm is now styled LLM in the documentation. #45

  • Examples in documentation now include a copy button. #43

  • llm templates command no longer has its display disrupted by newlines. #42

  • llm templates command now includes system prompt, if set. #44

0.4 (2023-06-17)#

This release includes some backwards-incompatible changes:

  • The -4 option for GPT-4 is now -m 4.

  • The --code option has been removed.

  • The -s option has been removed as streaming is now the default. Use --no-stream to opt out of streaming.

Prompt templates#

Templates is a new feature that allows prompts to be saved as templates and re-used with different variables.

Templates can be created using the llm templates edit command:

llm templates edit summarize

Templates are YAML - the following template defines summarization using a system prompt:

system: Summarize this text

The template can then be executed like this:

cat myfile.txt | llm -t summarize

Templates can include both system prompts, regular prompts and indicate the model they should use. They can reference variables such as $input for content piped to the tool, or other variables that are passed using the new -p/--param option.

This example adds a voice parameter:

system: Summarize this text in the voice of $voice

Then to run it (via strip-tags to remove HTML tags from the input):

curl -s 'https://til.simonwillison.net/macos/imovie-slides-and-audio' | \
  strip-tags -m | llm -t summarize -p voice GlaDOS

Example output:

My previous test subject seemed to have learned something new about iMovie. They exported keynote slides as individual images […] Quite impressive for a human.

The Templates documentation provides more detailed examples.

Continue previous chat#

You can now use llm to continue a previous conversation with the OpenAI chat models (gpt-3.5-turbo and gpt-4). This will include your previous prompts and responses in the prompt sent to the API, allowing the model to continue within the same context.

Use the new -c/--continue option to continue from the previous message thread:

llm "Pretend to be a witty gerbil, say hi briefly"

Greetings, dear human! I am a clever gerbil, ready to entertain you with my quick wit and endless energy.

llm "What do you think of snacks?" -c

Oh, how I adore snacks, dear human! Crunchy carrot sticks, sweet apple slices, and chewy yogurt drops are some of my favorite treats. I could nibble on them all day long!

The -c option will continue from the most recent logged message.

To continue a different chat, pass an integer ID to the --chat option. This should be the ID of a previously logged message. You can find these IDs using the llm logs command.

Thanks Amjith Ramanujam for contributing to this feature. #6

New mechanism for storing API keys#

API keys for language models such as those by OpenAI can now be saved using the new llm keys family of commands.

To set the default key to be used for the OpenAI APIs, run this:

llm keys set openai

Then paste in your API key.

Keys can also be passed using the new --key command line option - this can be a full key or the alias of a key that has been previously stored.

See API key management for more. #13

New location for the logs.db database#

The logs.db database that stores a history of executed prompts no longer lives at ~/.llm/log.db - it can now be found in a location that better fits the host operating system, which can be seen using:

llm logs path

On macOS this is ~/Library/Application Support/io.datasette.llm/logs.db.

To open that database using Datasette, run this:

datasette "$(llm logs path)"

You can upgrade your existing installation by copying your database to the new location like this:

cp ~/.llm/log.db "$(llm logs path)"
rm -rf ~/.llm # To tidy up the now obsolete directory

The database schema has changed, and will be updated automatically the first time you run the command.

That schema is included in the documentation. #35

Other changes#

  • New llm logs --truncate option (shortcut -t) which truncates the displayed prompts to make the log output easier to read. #16

  • Documentation now spans multiple pages and lives at https://llm.datasette.io/ #21

  • Default llm chatgpt command has been renamed to llm prompt. #17

  • Removed --code option in favour of new prompt templates mechanism. #24

  • Responses are now streamed by default, if the model supports streaming. The -s/--stream option has been removed. A new --no-stream option can be used to opt-out of streaming. #25

  • The -4/--gpt4 option has been removed in favour of -m 4 or -m gpt4, using a new mechanism that allows models to have additional short names.

  • The new gpt-3.5-turbo-16k model with a 16,000 token context length can now also be accessed using -m chatgpt-16k or -m 3.5-16k. Thanks, Benjamin Kirkbride. #37

  • Improved display of error messages from OpenAI. #15

0.3 (2023-05-17)#

  • llm logs command for browsing logs of previously executed completions. #3

  • llm "Python code to output factorial 10" --code option which sets a system prompt designed to encourage code to be output without any additional explanatory text. #5

  • Tool can now accept a prompt piped directly to standard input. #11

0.2 (2023-04-01)#

  • If a SQLite database exists in ~/.llm/log.db all prompts and responses are logged to that file. The llm init-db command can be used to create this file. #2

0.1 (2023-04-01)#

  • Initial prototype release. #1