Unlock the Full Power of LM Studio for Local LLM Deployment

Version 0.4 – Architectural Changes

llmster daemon : a head‑less process that separates the GUI from the inference engine, enabling deployment on machines without a graphical interface (cloud servers, GPU rigs, CI/CD pipelines, Google Colab). Installation is a single command and the daemon can be started, models downloaded, and an API server launched via CLI commands.

Parallel requests + continuous batching : based on llama.cpp 2.0, LM Studio now supports multiple concurrent inference requests. New model‑loading options are Max Concurrent Predictions (default 4) and Unified KV Cache , which share hardware resources with minimal memory overhead.

Stateful REST API : the /v1/chat endpoint returns a response_id and expects a previous_response_id to continue a conversation, reducing request payload size and providing token statistics, speed data, and permission‑key access.

UI refresh : added chat export (PDF/Markdown), split‑screen view, Developer Mode, and in‑app documentation.

Developer‑Facing Features

OpenAI‑compatible API – switch to a local model by changing the base URL

from openai import OpenAI
client = OpenAI(base_url="http://localhost:1234/v1")
response = client.chat.completions.create(
    model="<em>model‑id‑from‑LM Studio</em>",
    messages=[{"role": "user", "content": "Say this is a test!"}],
    temperature=0.7
)

Equivalent TypeScript and cURL examples work the same way, allowing local testing of agents, RAG pipelines, or AI workflows.

Anthropic‑compatible API – run Claude Code without an Anthropic API key

From version 0.4.1 LM Studio provides an /v1/messages endpoint compatible with Anthropic.

lms server start --port 1234
export ANTHROPIC_BASE_URL=http://localhost:1234
export ANTHROPIC_AUTH_TOKEN=lmstudio
claude --model openai/gpt-oss-20b

Python SDK example:

from anthropic import Anthropic
client = Anthropic(base_url="http://localhost:1234", api_key="lmstudio")
message = client.messages.create(
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello from LM Studio"}],
    model="ibm/granite-4-micro"
)
print(message.content)

CLI tool ( lms )

Install the CLI (bundled with llmster) and use the following commands:

npx lmstudio install-cli
lms status                # check LM Studio status
lms daemon up            # start the daemon
lms get <model>           # download a model
lms server start          # launch the API server
lms load <model>          # load a model into memory
lms chat                  # interactive terminal chat
lms ls --json            # list models in JSON (script‑friendly)
lms runtime update llama.cpp  # update the inference engine

The lms chat command supports slash commands such as /model, /download, /system-prompt, /help, and /exit, enabling a fully terminal‑based workflow: download → load → chat → debug.

Native SDKs

TypeScript SDK:

npm install @lmstudio/sdk
import { LMStudioClient } from "@lmstudio/sdk"
const client = new LMStudioClient()
const model = await client.llm.model("openai/gpt-oss-20b")
const result = await model.respond("Who are you, and what can you do?")
console.info(result.content)

Python SDK:

pip install lmstudio
import lmstudio as lms
with lms.Client() as client:
    model = client.llm.model("openai/gpt-oss-20b")
    result = model.respond("Who are you, and what can you do?")
    print(result)

The SDKs expose advanced capabilities: tool calling, MCP integration, structured JSON output, embeddings, tokenization, and full model management (download, load, list, unload).

LM Link – remote model loading via Tailscale mesh VPN

LM Link creates a secure, end‑to‑end encrypted tunnel between multiple devices (e.g., a home 4090 server and a work laptop). The local localhost:1234 endpoint forwards requests to the remote GPU machine while keeping chat data local.

Based on Tailscale mesh VPN; no public ports are exposed.

Chat data stays on the client; inference runs on the remote device.

The same localhost:1234 API works for Codex, Claude Code, OpenCode, etc.

Free tier: 2 users, up to 10 devices (5 per user).