1.6.0b1 - 2025-10-28

What's Changed

• Handle pre-formatted server API errors by @ncoghlan in #165
• API token support by @ryan-the-crayon and @ncoghlan in #163

New Contributors

• @ryan-the-crayon made their first contribution in #163

Full Changelog: 1.5.0...1.6.0b1

1.5.0 - 2025-08-22

What's Changed

• Accept any buffer instance when processing file data (by @ArEnSc; #46)
• Tool definitions now support default parameter values (by @baonudesifeizhai; #90)
• Preliminary addition of Python 3.14 to the CI test matrix (#154). Most features should work without issue, but there is (for now) a known problem with field definitions on lmstudio.BaseModel subclasses only being processed correctly if from __future__ import annotations is correct at the time the subclass is defined.

New Contributors

• @ArEnSc made their first contribution in #148
• @baonudesifeizhai made their first contribution in #150

Full Changelog: 1.5.0b1...1.5.0

1.5.0b1

What's Changed

• When no API host is specified, the SDK now attempts to connect to the always-on LM Studio native API server ports, rather than relying on the optional HTTP REST API server being enabled on its default port (#142)
• The new methods Client.is_valid_api_host and AsyncClient.is_valid_api_host allow API host validity to be checked without creating a client instance (part of #142)
• The new methods Client.find_default_local_api_host and AsyncClient.find_default_local_api_host allow discovery of a running local API host API server without creating a client instance (part of #142)
• Exceptions are now more consistently raised on websocket failures (#121). Previously clients could potentially be left hanging indefinitely if the websocket connection failed while waiting for a response.
• Updated to 2025-07-30 (release 54) lmstudio.js protocol schema (#138)
• The low level session APIs are now more explicitly private (in both the synchronous and asynchronous APIs)

Changes specific to the asynchronous API

• The asynchronous API is now considered stable and no longer emits FutureWarning when imported (#127 and supporting PRs)
• Asynchronous model handles now provide a model.act() API (#132)
• The asynchronous model.act() API supports asynchronous tool implementations (#134)

Changes specific to the synchronous API

• The synchronous API now implements a configurable timeout (#124). The default API timeout if no messages are received from the server on a given response channel is now 60 seconds.
• The new functions get_sync_api_timeout() and set_sync_api_timeout(timeout: float|None) allow this timeout to be queried and updated (part of #124). Setting the timeout to None restores the previous behaviour of blocking indefinitely waiting for server responses.
• The synchronous API now supports invocation from atexit hooks (#123, #125)

Full Changelog: 1.4.1...1.5.0b1

1.4.1

What's Changed

• Fix handling of multi-part tool results in Chat.append and Chat.add_entry (#112)
• Fix server mapping for llama quantization config settings (#111)
• More lenient handling of received config fields with omitted default values (#110)

Full Changelog: 1.4.0...1.4.1

1.4.0

What's Changed

• Add max_parallel_tool_calls to .act() API, defaulting to sequential execution (#102)
• Unknown channel messages are now reported via warnings.warn rather than immediately raising an exception (#103)

Full Changelog: 1.3.2...1.4.0

1.3.2

What's Changed

• Update to 2025-06-25 lmstudio.js protocol schema (#99)

Full Changelog: 1.3.1...1.3.2

1.3.1

What's Changed

• Fixed GPU offload configuration when loading models (#89)

Full Changelog: 1.3.0...1.3.1

1.3.0

What's Changed

• Added a dedicated configure_default_client API (#82)
• Runtime names of public API types now match their import names (#74)
• Tool call failures are now passed back to the LLM by default (#72)
• Synchronous clients now use fewer background threads (#77)
• Tool results now report non-ASCII characters as Unicode code points rather than as ASCII escape sequences (contributed by @jusallcaz in #80)
• The not-yet-implemented general file handling APIs have been removed from the public API (#81)

New Contributors

• @jusallcaz made their first contribution in #80

Full Changelog: 1.2.0...1.3.0

1.2.0 - 2025-03-22

What's Changed

• Align model loading config format with lmstudio-js 47 (previously aligned with lmstudio-js 46)
  • Note: users of the experimental gpuOffload config setting will need to switch to setting the (still experimental) gpu field
• Explicitly note in relevant docstrings that details of config fields are not yet formally stabilised
• Pass previously omitted config settings to the server (#51)
• Add speculative decoding example (#55)
• Publish config retrieval APIs (#53, #54)
• Add server side token counting APIs (#57)
• Simplify prediction API type hinting (#59)
• Add preset config support in prediction APIs (#58) (Requires LM Studio 0.14+)
• Add GBNF grammar support when requesting structured responses (#60) (Requires LM Studio 0.14+)

Full Changelog: 1.1.0...1.2.0

1.1.0 - 2025-03-15

What's Changed

• Added SDK versioning policy to README.md (#42)
• Support Python 3.10 (#41)
• Support image input for VLMs (#34)
• Publish file preparation APIs (#37, #39)
• Add image preparation APIs (#38, #39)
• Avoid corrupting snake_case keys in structured output schemas supplied via prediction config dicts (#43)

Full Changelog: 1.0.1...1.1.0