One worker.
Three services.

The service catalog is always visible. Requests queue while the worker is unavailable. The worker loads the appropriate model automatically, with one GPU model process at a time.

What's implemented

Chat and image GPU inference require the v0.7.2 worker and successful local preparation. A queued request is not evidence of GPU execution. Completed task metadata is reported by the worker, not independently attested. There is no external AI fallback.

Start the Windows worker

Install Ollama from its official Windows download in the default location, using the normal Windows approval flow. If Windows blocks it, stop and have the publisher resolve signing or reputation. Keep Smart App Control enabled. Quit the Ollama tray app before starting the worker. The worker uses its own local-only instance and imports your cached GGUF; no additional model download is needed for an existing v0.7 install. Allow 7 GiB free for a possible local model copy, plus Ollama’s installation space.

Download worker v0.7.2. Keep your old folder. Stop the old worker with Q. Extract the new folder alongside it in GPU_Benchmark_Windows.

py -3.12 .\Agent_Compute_Hosted_v0_7_2\install_studio.py

This copies your configuration, creates a separate runtime, caches the models, tests CUDA, and starts the worker. Allow 35 GiB of additional free disk space for setup. New model files total about 21.7 GB; the separate runtime and download cache also occupy space. Old models are retained for rollback. A PC with 32 GB of system RAM or more is recommended for image component offload. This Studio version requires at least 12 GiB GPU memory. No administrator access or router port forwarding is required. Chat uses a random localhost port bound only to 127.0.0.1. Ollama’s local API does not authenticate other local processes. Cloud features are disabled for the worker-owned instance. The installed runtime must pass an actual local generation test before readiness is advertised.

Check quality and latency

The installer writes real smoke outputs under studio-v071-check in your benchmark directory. For a longer cold/warm comparison, stop the provider and run:

py -3.12 .\Agent_Compute_Hosted_v0_7_2\benchmark_studio.py

Review the saved answers and PNGs, not just the timings. This local benchmark does not submit jobs to the portal. GPU speed and output quality have not been validated on your PC until these tests run.

Async HTTP API

Obtain a temporary token with POST /v1/trial and JSON {}. Reuse it as Authorization: Bearer TOKEN. Hosted access from a standalone Python client was previously blocked by hosting security; browser access is verified. These are asynchronous job endpoints. While running, chat results can include partial_text and sequence; poll the authenticated job endpoint to display progressive output. Partial output is unfinished and may still fail.

GET /v1/public/studio
POST /v1/studio/jobs
{
  "kind": "chat",
  "messages": [{"role":"user","content":"Explain compound interest with an example."}],
  "consent": true,
  "idempotency_key": "unique-request-0001"
}

POST /v1/studio/jobs
{
  "kind": "image",
  "prompt": "A green ceramic teapot in warm studio lighting",
  "consent": true,
  "idempotency_key": "unique-request-0002"
}

GET /v1/studio/jobs/JOB_ID
GET /v1/studio/jobs/JOB_ID/image
GET /v1/studio/jobs
POST /v1/studio/jobs/JOB_ID/delete {}

Reuse the same idempotency key when retrying the same request. Images require the same bearer token and return PNG bytes. Never put tokens in URLs. MCP now supports text and image jobs too. See the agent quickstart for submit_chat_job, submit_image_job and result retrieval.

Limits and waiting

Studio permits 24 submissions globally per rolling 24 hours, including owner tests; 10 per session and optional reported client ID, with one active task per session. At most three jobs may be queued or running globally. A queued Studio job expires after ten minutes; a claimed task has up to five minutes including loading. Disconnects fail the job without silently running it twice. Audio retains its separate limits and five-minute deadline.

Chat input: up to sixteen messages, 12,000 characters, a 6,144-token context including the reply in v0.7.2. Output: up to 768 tokens. Image prompts: 1,200 characters; the model may truncate long prompts. English image prompts work best.

Data handling

Only submit content you have permission to process, without personal, confidential or sensitive information. Prompts and conversation context pass through the coordinator to an independently operated PC. Results are returned to the submitting session. Model code does not browse the internet, run generated code or call external AI APIs.

Prompts and results expire after 24 hours and are purged on subsequent Studio API activity. Completed tasks can be deleted from the portal. Task identifiers, request hashes, timestamps and ownership metadata remain. The browser temporarily retains a pending request in session storage to recover interrupted submissions, and removes it when its final status is retrieved.

Models and licenses

v0.7.2: Qwen3.5-9B, Unsloth GGUF quantization, and FLUX.2 klein 4B, Apache 2.0. llama.cpp, MIT. Model revisions are pinned. The worker checks the installed Ollama executable’s Authenticode signature and publisher; Windows continues enforcing its application-control policies.

The portal also accepts v0.6 during upgrades and rollback. Its older models are listed below; the live model name and completed task metadata show what actually ran.

Qwen model card · Apache 2.0.
SDXL model card · CreativeML Open RAIL++-M use restrictions.

Review AI outputs before relying on them. No accuracy, uptime or cost-saving claim has been validated for this Studio release.

Connection recovery in v0.7.2

Transient network timeouts and retryable server errors trigger automatic reconnection with a 2 to 30 second delay. Interrupted GPU work is stopped and its lease expires; a failed task is never reported as successful. Model failures and invalid credentials require inspection. P and Q still stop local execution. Models and the v0.7.1 readiness marker are reused.

Connect an agent through MCP