Machine Learning/LiftWing/Large Language Models/Wikimania 2026

This page is currently a draft.
Material may not yet be complete, information may presently be omitted, and certain parts of the content may be subject to radical, rapid alteration. More information pertaining to this may be available on the talk page.

Part of LiftWing Large Language Models · Wikimania 2026

LiftWing hosts open-weight LLMs you can use in your Wikimedia projects and tools. This page gets you from zero to a working request in a few minutes — chat with a model in your browser, or call it from your code. For the full platform reference (internal models, deployment, performance), see the main LLM documentation.

Available models

Two general-purpose Qwen chat models are open for you to use:

Model	Model card	Context	Notes
`llm-qwen3-14b`	Qwen3-14B	16K tokens	14B general-purpose chat model. A good default.
`llm-qwen36-27b`	Qwen3.6-27B	32K tokens	27B general-purpose chat model; the largest available.

Both are multilingual, instruction-tuned chat models.

Responses stream token-by-token, so output appears gradually — expect a short wait for the first token. The service is shared, so it may be slower under heavy load. For measured throughput and latency, see Performance in the main documentation.

→ More detail: the full model list (including internal models) and per-model specs in Available Models and Deployment Information in the main documentation.

Quick start

The fastest ways to try a model:

No code: open LiftWing Studio and chat with a model in your browser.
From the terminal: the public API is OpenAI-compatible, so a single curl gets you a completion.

curl https://api.wikimedia.org/service/lw/inference/v1/models/llm-qwen3-14b/openai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "llm-qwen3-14b", "messages": [{"role": "user", "content": "In two sentences, what is the Wikimedia Foundation?"}]}'

No API key is required. Anonymous access is shared across everyone at 100 requests/hour — for anything beyond quick testing, run it from Toolforge for a much higher limit.

LiftWing Studio

LiftWing Studio gives the models a friendly chat interface, so you can use them in your browser instead of through the API — pick a model, type a prompt, and iterate. It's the recommended first stop for exploring what the models can do.

Access

Register an account on the Studio sign-in page. New accounts are activated after a quick admin approval.

What it offers

Chat history across sessions
Document uploads
Model selection from the available models

→ More detail: how Studio is built, in LiftWing Studio in the main documentation.

API usage

Public models are served through the Wikimedia REST Gateway at api.wikimedia.org with an OpenAI-compatible chat-completions interface, so existing OpenAI client libraries work by pointing base_url at the model's endpoint. No API key is required.

Endpoint:

https://api.wikimedia.org/service/lw/inference/v1/models/llm-<model>/openai/v1/chat/completions

You can do the same from Python with any OpenAI client library:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.wikimedia.org/service/lw/inference/v1/models/llm-qwen3-14b/openai/v1",
    api_key="none",  # public endpoint; no key required
)

resp = client.chat.completions.create(
    model="llm-qwen3-14b",
    messages=[{"role": "user", "content": "Explain vLLM in one sentence."}],
    stream=True,   # responses stream token-by-token
)
for chunk in resp:
    print(chunk.choices[0].delta.content or "", end="")

The models may include their reasoning wrapped in <think>…</think> at the start of a response — strip it if you only want the final answer.

For the full request/response schema, see the OpenAI Chat Completions reference.

Higher rate limit from Toolforge

Running from Toolforge lifts the shared 100 req/hour cap to effectively unlimited — ideal for a hackathon tool. For a quick call you don't need to build anything:

Create a Wikimedia developer account and join Toolforge — see the Toolforge Quickstart.
Log in to the bastion: ssh <your-username>@login.toolforge.org
Run the same curl as above — from Toolforge it automatically gets the higher limit.

Building a tool or application? The bastion is for quick, interactive use. For anything sustained — a running service, scheduled jobs, or a shared project — create a dedicated tool account and run it as a job or web service.

→ More detail: the full rate-limit tiers and access-by-platform in Rate Limits & Access in the main documentation.

What it can and can't do

Works today	Not available yet
Chat & text completions (OpenAI-compatible)	Tool / function calling
Streaming responses	Web search / browsing
Context window up to 32K tokens	Retrieval-augmented generation (RAG)
Many languages	Image / multimodal input

FAQ

FAQ questions and answers to be added here — TBD.

Get help

Report an issue: Phabricator, tagged Machine-Learning-Team
Chat: IRC #wikimedia-ml (Libera Chat)
Email: ml@wikimedia.org

For full platform details, see the main LLM documentation.