GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Fluxmire
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Fluxmire
Model Tree
computer-vision-ai-lab/Qwen-3.6-27B-AmberKite
Input Modalities
Output Modalities
Supported Functionality
| Architecture | Qwen3_5ForConditionalGeneration — 64 decoder layers (hybrid linear / full attention), hidden size 5120, 27-block vision encoder |
| Context length | 262 144 tokens |
| Weights | language-model Linear layers in FP8 (compressed-tensors, dynamic per-token activation scales); vision tower, linear-attention blocks, lm_head and MTP head in bfloat16 |
| Files | model-00001-of-00002.safetensors, model-00002-of-00002.safetensors (~31 GB), model-mtp.safetensors (optional multi-token-prediction head), tokenizer, processor and chat template |
| Quantization record | recipe.yaml |
Tokenizer, processor configuration and chat template are those of the Qwen3.6 family and are unchanged.
Requires vLLM ≥ 0.23 with compressed-tensors FP8 support.
bash
vllm serve magicworld7/Qwen-3.6-27B-Merlin \ --revision <commit> \ --data-parallel-size 2 \ --max-model-len 81920 \ --gpu-memory-utilization 0.9 \ --enable-prefix-caching --enable-chunked-prefillThe server exposes the OpenAI-compatible chat API; image inputs are passed as image_url parts
of the user message. generation_config.json holds the family's default sampling parameters
(temperature 1.0, top_p 0.95, top_k 20); adjust them per use case.
Pin a revision when reproducibility matters — the repository may receive further commits.
python
from transformers import AutoProcessor, AutoModelForImageTextToTextm = "magicworld7/Qwen-3.6-27B-Merlin"processor = AutoProcessor.from_pretrained(m)model = AutoModelForImageTextToText.from_pretrained(m, device_map="auto")A recent Transformers release with compressed-tensors installed is needed to read the FP8
weights.
Apache License 2.0. LICENSE contains the license text; NOTICE contains the change notice and
upstream attribution required by Section 4 of the license. "Qwen" is a trademark of Alibaba Cloud
and is used only to identify the model family.
vllm serve magicworld7/Qwen-3.6-27B-Merlin \ --revision <commit> \ --data-parallel-size 2 \ --max-model-len 81920 \ --gpu-memory-utilization 0.9 \ --enable-prefix-caching --enable-chunked-prefillfrom transformers import AutoProcessor, AutoModelForImageTextToTextm = "magicworld7/Qwen-3.6-27B-Merlin"processor = AutoProcessor.from_pretrained(m)model = AutoModelForImageTextToText.from_pretrained(m, device_map="auto")