← Back to Blog

Use OpenAI's Responses API with local models

By
LM Studio Team

•

2025-10-06

LM Studio 0.3.29 is now available as a stable release. Update in‑app or download the latest version at lmstudio.ai/download.

New: OpenAI-compatible /v1/responses

This release adds support for OpenAI’s /v1/responses API through the LM Studio REST server.

  • Stateful interactions - pass a previous_response_id to continue interactions without needing to manage message history yourself.
  • Custom function tool calling - bring your own function tools for the model to call, similar to in v1/chat/completions.
  • Remote MCP - enable the model to call tools from remote MCP servers, requires explicit opt-in in settings.
  • Reasoning support - parse reasoning output, and control effort with reasoning: { effort: "low" | "medium" | "high" } for openai/gpt-oss-20b.
  • Streaming or sync - use stream: true to receive SSE events as the model generates, or omit for a single JSON response.
  • More information at https://lmstudio.ai/docs/app/api/endpoints/openai

To use REST API server endpoints, ensure your LM Studio server is running in the UI (Developer → Status: Running):

undefined

Start the server in the UI

Or through the lms CLI:

→ % lms server start
Success! Server is now running on port 1234

Stateful Interactions

  • Example request:

    curl http://127.0.0.1:1234/v1/responses \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-oss-20b",
    "input": "Provide a prime number less than 50",
    "reasoning": { "effort": "low" }
  }'

  • Response (shortened):

    { "id": "resp_123", "output": [{"type":"message", ...}] }

  • Continue the interaction by setting previous_response_id to the id above:

    curl http://127.0.0.1:1234/v1/responses \
-H "Content-Type: application/json" \
-d '{
    "model": "openai/gpt-oss-20b",
    "input": "Multiply it by 2",
    "previous_response_id": "resp_123"
}'

Streaming

Example request:

curl http://127.0.0.1:1234/v1/responses \
  -H "Content-Type: application/json" \
    -d '{
      "model": "openai/gpt-oss-20b",
      "input": "Hello",
      "stream": true
    }'

You’ll receive events as the model generates output like response.created, response.output_text.delta, and response.completed. See docs for more details on streaming events.

Remote MCP

Opt-in to allow use of remote MCP servers (Developer → Settings → Allow MCP → Remote):

undefined

Allow remote MCP servers

Example request:

curl http://127.0.0.1:1234/v1/responses \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -d '{
    "model": "openai/gpt-oss-20b",
    "tools": [{
      "type": "mcp",
      "server_label": "tiktoken",
      "server_url": "https://gitmcp.io/openai/tiktoken",
      "allowed_tools": ["fetch_tiktoken_documentation"]
    }],
    "input": "What is the first sentence of the tiktoken documentation?"
  }'

Output will include a tool discovery and a tool call before the assistant’s reply. See docs for full schema and examples.

CLI: list all model variants

Quickly inspect every available variant for multi‑variant models:

lms ls --variants

Example output (variants only):

google/gemma-3-12b (2 variants)
 * google/gemma-3-12b@q3_k_l   12B   gemma3   7.33 GB
   google/gemma-3-12b@4bit     12B   gemma3   8.07 GB

0.3.29 - Release Notes

Build 1

  • New OpenAI compatibility endpoint: /v1/responses
    • Create stateful interactions by passing the id of a previous response as input to the next — no need to manage message history yourself
    • Custom tool calling support
    • Supports reasoning parsing, setting reasoning effort ("low"|"medium"|"high") for models like gpt-oss
    • Synchronous and streaming (stream=true)
    • See https://lmstudio.ai/docs/app/api/endpoints/openai for more details
  • New lms ls command option: lms ls --variants to list all variants for multi-variant models

Resources