Skip to content
Computer vision

Qwen3.81-27B

Merging Two Qwen3 Fine-Tunes

1 min readupdated

The model is a 50/50 linear parameter average of two Qwen3.5-based versions: the agent-terminal DaisyRose-27B and the multimodal sabi-noema with hybrid attention. Thanks to the orthogonality of their fine-tuning directions, the merge retains about 100% of the combined signal without losing knowledge. The model weighs 54.7 GB in bf16 format and is well-suited for programming, reasoning, and image tasks.

Key features

  • Orthogonal merge: The fine-tuning directions of the two base models are perpendicular, allowing their experience to be combined successfully.
  • Multimodality: Fully preserves the image processing architecture from the base Qwen3.5 lineup.
  • Reasoning mode: Supports a chat template with a built-in deep thinking format.
  • Quantization option: An NVFP4 version (19 GB) is available to run on a single GPU.

What you can do

  • Run via vLLM: Deploy a local server to work with the model and long context.
  • Solve terminal tasks: Use the strong agent skills inherited from DaisyRose.
  • Multimodal analysis: Process queries combining text and images thanks to the hybrid architecture.

Example

vllm serve vince62s/Qwen3.81-27B \
  --served-model-name qwen3.81 \
  --reasoning-parser qwen3 \
  --max-model-len 32768
Read moreDetails
• GPQA Diamond: 43.9% (loglikelihood)
• VRAM requirements: ~55 GB+ for the bf16 version
• Parameters: 27.36 billion parameters, safetensors format (3 shards)
• Sampling: temperature 1.0, top_p 0.95
Discuss Telegram 0 21

Will it run?

Runs in 4-bit

Full precision does not fit, a quantised version does: quality barely changes down to 8 bits, more noticeably at 4 bits.

  • BF16weights 51 GB · ready files57 GB
  • FP8 / INT8recommendedweights 25 GB29 GB
  • 4-bitweights 14 GB16 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # vince62s/Qwen3.81-27B
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download vince62s/Qwen3.81-27B --local-dir ~/models/Qwen3.81-27B
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to Qwen3.81-27B
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

  1. Qwen3.8-27Bfamily root
  2. DaisyRose-27B
  3. Qwen3.81-27B

New posts about versions of Qwen3.8-27B (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 2All on Hugging Face

About the model

vince62s/Qwen3.81-27B
Author
vince62s
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

210

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments

Language models

sid1-qwen3-4b

sid1-qwen3-4b