Skip to main content

Qwen3.8 (Open-Weight)

Qwen3.8 is Alibaba's open-weight multimodal family on Hugging Face — the open counterpart to the hosted Qwen3.8-Max DashScope API. Use this for image/video understanding on your own GPUs (privacy, no per-request cost, offline).

Capabilities​

Same series strengths as hosted Max, at open-weight scale:

CapabilityOpenTryOn local (qwen3.8)
ModalitiesText + image (+ video via frame sampling) → text
Default weightsDense Qwen3.8-27B (~27B VLM)
Context262K native; toward 1M with YaRN in vLLM/SGLang
ThinkingOn by default; --no-thinking / enable_thinking=False
Coding / agents / toolsStrong in the model family; local adapter focuses on understand (use a serving stack for full agent loops)
Cluster Max-classQwen/Qwen3.8-2.4T-A95B MoE — not single-GPU Transformers

Useful for garments, lookbooks, documents, UI screenshots, product photos, and sampled video — the same general-purpose understand path as the API adapter.

Overview​

FeatureValue
Default modelQwen/Qwen3.8-27B
ArchitectureDense VLM (~27B), hybrid attention + vision encoder
Context length262K native (extensible toward 1M with YaRN in serving stacks)
OpenTryOn focusImage, multi-image, and frame-sampled video understanding; thinking mode
Min VRAM~50GB+ recommended for bf16; community quantized checkpoints for smaller cards

The flagship open MoE Qwen/Qwen3.8-2.4T-A95B is datacenter-scale — pass model_id="Qwen/Qwen3.8-2.4T-A95B" only with cluster serving (vLLM / SGLang), not single-GPU Transformers.

Installation​

pip install opentryon[local]
pip install -U transformers   # Qwen3.8 needs a recent transformers
pip install decord            # only for understand_video()

Quick Start​

Image Understanding​

from tryon.models import Qwen38Adapter

adapter = Qwen38Adapter()  # downloads Qwen/Qwen3.8-27B on first use

result = adapter.understand_image(
    "garment.jpg",
    prompt="Describe this outfit: color, pattern, style, fit, and material.",
)
print(result["text"])
print(result["thinking"])

Video Understanding​

Video uses uniform frame sampling as a multi-image prompt (transformers fallback; native video IO is better under vLLM/SGLang):

result = adapter.understand_video(
    "runway_clip.mp4",
    prompt="Summarize the styling shown in this video.",
    num_frames=8,
)
print(result["text"])

API Reference​

Qwen38Adapter​

class Qwen38Adapter:
    def __init__(
        self,
        model_id: Optional[str] = None,     # or QWEN38_MODEL_ID / Qwen/Qwen3.8-27B
        device: Optional[str] = None,       # device_map, default "auto"
        torch_dtype: str = "auto",
        trust_remote_code: bool = True,
    )

Methods​

  • understand_image(image, prompt, max_new_tokens=4096, temperature=0.8, enable_thinking=True)
  • understand_video(video, prompt, num_frames=8, ...)
  • understand(image=None, video=None, ...) — CLI / MCP entry point

Returns {"thinking": ..., "text": ..., "model": ...}.

CLI​

opentryon understand --model qwen3.8 --image garment.jpg \
  --prompt "Describe this outfit."

opentryon understand --model qwen3.8 --video runway_clip.mp4 --num-frames 12

opentryon understand --model qwen3.8 --image garment.jpg --no-thinking

MCP​

Local registry id qwen3.8 → tool understand_qwen3_8 (requires pip install opentryon[local]). Hosted Max is understand_qwen3_8_max. See MCP Server.

Choosing a Variant​

ModelScaleUse case
Qwen/Qwen3.8-27B (default)Dense 27BPractical open multimodal reasoning
Qwen/Qwen3.8-2.4T-A95BMoE ~2.4T / 95B activeCluster-scale; closest to hosted Max
Community AWQ/NVFP4 quantsSmaller VRAMSingle-GPU when bf16 does not fit
adapter = Qwen38Adapter(model_id="barrydeen/Qwen3.8-27B-AWQ-4bit")

Troubleshooting​

Out of Memory​

  • Prefer a quantized Hub checkpoint, or serve 27B with vLLM/SGLang.
  • Do not load Qwen3.8-2.4T-A95B on a single GPU.

AutoModelForImageTextToText / architecture errors​

  • Upgrade transformers: pip install -U transformers.

ImportError for decord​

  • Needed only for video frame sampling: pip install decord.

References​