Our inference API can now see. As of today we serve Qwen 3.8 27B — Alibaba's newest open model, and the first in the family that is natively multimodal: there is no separate "VL" variant, the base model itself takes text and images in. It's live on the ParalonCloud API right now, OpenAI-compatible, with tool calling and a 32k context window. And like everything on our API during the beta: free, 60 requests/minute, no card required.
Model id: qwen3.8-27b. If you were using qwen3.6-27b, that model is retired — swap the id and you're migrated (same context, same tool-call parser, plus eyes).
Send it an image
The endpoint is the same Chat Completions API you already use — images ride along as standard OpenAI-style content parts. This exact request ran against production before we published:
curl https://paraloncloud.com/v1/chat/completions \
-H "Authorization: Bearer prlc_your_api_key" \
-H "Content-Type: application/json" \
-d '{
"model": "qwen3.8-27b",
"messages": [{
"role": "user",
"content": [
{"type": "text", "text": "What solid color is this image?"},
{"type": "image_url", "image_url": {"url": "data:image/png;base64,iVBORw0K..."}}
]
}]
}'
"The image is a uniform field of pure red (RGB approximately 255,0,0)."
Base64 data-URLs and regular https:// image URLs both work. With the OpenAI Python SDK it's the usual two-line change — point base_url at https://paraloncloud.com/v1 and pass your prlc_ key. No key yet? Sign in to the Console, create one, done — the authentication docs have the details.
Want to try it without writing code? The Playground runs the same model in the browser.
The interesting part: your coding agent can read screenshots now
Text-only APIs have always had a blind spot in agentic coding: the agent can edit your CSS but can't see that the button is now overlapping the navbar. A vision model behind the same endpoint closes that loop — paste a screenshot, the agent looks at it, then fixes the code.
One thing to know: coding agents decide client-side what a model can do. If the model's config entry says text-only, the agent will refuse to attach images before a single byte reaches the API. For pi (full setup in our pi guide), that's the input array in ~/.pi/agent/models.json:
{
"id": "qwen3.8-27b",
"name": "Qwen 3.8 27B (Vision)",
"input": ["text", "image"],
"contextWindow": 32768,
"maxTokens": 8192
}
Other agents (OpenClaw, Cline, Roo Code, Continue…) have an equivalent capability flag on custom models. If your agent says "model does not support images" while pointing at qwen3.8-27b, it's that flag — the API side is fine.
What's under the hood
Qwen 3.8 27B uses a hybrid attention architecture (48 linear-attention layers, 16 full-attention) which makes its KV cache dramatically cheaper than a conventional 27B — that's what lets us serve a multimodal model of this class on single 24GB consumer GPUs across our distributed fleet, quantized to W4A16 with the vision tower kept at full precision. If you're curious how we schedule models like this across a heterogeneous swarm of 3090s, 4090s and 5090s, we wrote about it in Scheduling LLMs across mixed GPUs.
Current serving limits, so you can build against them honestly:
| Context window | 32,768 tokens |
| Images per request | up to 2 |
| Formats | PNG, JPEG, WebP, GIF (as data-URL or https URL) |
| Tool calling | native (tool_calls), same parser as 3.6 |
| Price | free during beta · 60 req/min |
Why this matters beyond the demo
Vision-capable APIs are mostly locked behind closed vendors with per-image pricing. An open-weight multimodal model on an OpenAI-compatible endpoint means UI-debugging agents, OCR pipelines, document extraction, screenshot-driven test triage — all buildable today, at zero cost during beta, with no vendor account. The model list at /v1/models always reflects what's live.
Ship something with it and tell us on Discord — the fastest feedback loop we have.
