LiveMLX runtime
Chat and LLM inference
Stream chat completions through an OpenAI-compatible endpoint. Model availability is checked when you run a request.
Open chat →Open-model AI workloads, through one API.
OpenAI-compatible chat, vision Q&A, and fine-tuning. Requests are matched to a compatible provider device and quoted before they run. Live capacity is checked at request time.
Open alphaPython · TypeScript · RESTNo infrastructure to reserve
Preparing the request path…
REQUESTExplain unified memory in one sentence.qwen3-4b
RESPONSE
Unified memory lets the CPU and GPU work from the same pool of data…
01OpenAI-compatible
Keep the chat client you already use.
02Quoted first
See the unit rate and estimated total, then set a hard spend cap.
03Native runtimes
MLX workloads on compatible Mac computers with Apple silicon.
04One catalog
Discover workload, model, and lifecycle status.
Current catalog
Start with a model. Stay for the useful work around it.
Chat is the front door. The same account and API also handle vision Q&A and fine-tuning as your application grows.
PreviewQwen2-VL 2B
Vision Q&A
Ask questions about an image with a model selected from the live catalog.
Open vision Q&A →PreviewMLX training
LoRA fine-tuning
Create lightweight adapters for supported language models.
Open fine-tuning →PreviewReal-ESRGAN x4
Image upscale
Upscale an image to four times its width and height. You get a PNG back.
Open upscale →LiveWhisper large-v3-turbo
Audio transcription
Turn speech into timestamped text in 99+ languages, with the language detected for you.
Open transcription →LiveWhisper large-v3-turbo
Subtitles
Generate SRT or WebVTT subtitles timed to the speech in an audio or video file.
Open subtitles →Use the interface you know
Point an OpenAI client at Common Compute.
Change the base URL, choose an available model, and stream the response. The same account gives you the native SDK, CLI, MCP server, and generic job API.
chat.pyOpenAI-compatible
import os
from openai import OpenAI
client = OpenAI(
api_key=os.environ["CC_API_KEY"],
base_url="https://api.commoncompute.ai/v1",
)
stream = client.chat.completions.create(
model="qwen3-4b",
messages=[{"role": "user", "content": "Explain unified memory."}],
extra_body={
"data_class": "public",
"marketplace_execution_risk_acknowledged": True,
},
stream=True,
)
for event in stream:
print(event.choices[0].delta.content or "", end="")›Unified memory lets the CPU and GPU share one pool of data
Why Apple Silicon
One architecture. Several purpose-built AI paths.
Common Compute routes each workload to a Mac computer with the memory, operating system, and runtime it needs. The currently offered catalog uses MLX workloads and explicitly matched model requirements.
MLX
Language and generative models run through an Apple Silicon-native machine learning stack.
Open models
Choose from the models currently listed for chat, vision, and fine-tuning workloads.
Capability matching
The router checks workload, model, memory, and runtime requirements before dispatch.
Know how the network works
Marketplace compute, described plainly.
Jobs run on independently operated, provider-owned Mac computers. Common Compute records the selected device, price, usage, and receipt evidence so you can review what happened after a job completes.
Read the trust model01
Before: see a quote and choose a spend limit.
02
During: the router matches explicit capability requirements.
03
After: inspect job history, usage, and signed receipt evidence.
For Mac computer owners
Have a compatible Mac computer with Apple silicon?
Choose when it can accept work, which workloads it may run, and pause it at any time. Completed jobs appear in an itemized earnings ledger. New providers can connect a Mac computer now, but it does not receive paid work until the provider agreement review is complete.
Provider information →Signed and notarized macOS app