Skip to content
Computer vision

Qwen3.5-2B-NoRA

A model for reasoned action analysis from frames

2 min readupdated

The model is built on the Qwen/Qwen3.5-2B architecture (2.21 billion parameters) and fine-tuned on NoRA annotations for normative reasoning based on first-person visual scenes. Using preceding frames, the model describes observable facts, the underlying reasons derived from them, and potential subsequent actions with cross-links. This development is designed for research tasks in multimodal behavior and logic analysis. It will be useful for AI researchers and computer vision developers studying decision-making in visual environments.

Key features

  • Based on the efficient Qwen3.5-2B language model
  • Trained on GPT-5.4 annotations from the LLMSilver-1230 dataset
  • Full fine-tuning completed using the MS-SWIFT tool
  • Builds structured connections between facts, reasons, and actions

What you can do

  • Analyze first-person video sequences for research projects
  • Generate detailed structured JSON scene descriptions with action reasoning
  • Integrate the model into image and text processing pipelines via the transformers library

Example

import torch
from transformers import AutoModelForImageTextToText, AutoProcessor

name = "MINTLABJHUANU/Qwen3.5-2B-NoRA"
processor = AutoProcessor.from_pretrained(name)
model = AutoModelForImageTextToText.from_pretrained(
    name, torch_dtype=torch.bfloat16, device_map="auto")

SYSTEM_PROMPT = (
    "Describe a reasonable next-action decision from the camera wearer's "
    "perspective. Use only the visible scene. Return one JSON object with "
    "clip_id, facts, reasons, and actions. Return JSON only."
)
USER_PROMPT = "Create a NoRA annotation for clip_id `{clip_id}` from these pre-action frames."

messages = [
    {"role": "system", "content": [{"type": "text", "text": SYSTEM_PROMPT}]},
    {"role": "user", "content": [
        {"type": "image", "url": "frames.jpg"},
        {"type": "text", "text": USER_PROMPT.format(clip_id="example")},
    ]},
]
inputs = processor.apply_chat_template(
    messages, add_generation_prompt=True, tokenize=True,
    return_dict=True, return_tensors="pt", enable_thinking=False,
).to(model.device)
output = model.generate(**inputs, max_new_tokens=2048, do_sample=False)
print(processor.decode(
    output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
Read moreDetails
Training method: Full fine-tuning with MS-SWIFT
Epochs and learning rate: 3 / 1e-5
Batch size and gradient accumulation: 1 / 16
Precision and token limit: BF16 / 4,096 tokens
Limitations: The model is intended for normative reasoning research, is not a safety system, and must not be used for making decisions about people.
Discuss Telegram 0 37

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 4.1 GB · ready files5.0 GB
  • FP8 / INT8recommendedweights 2.1 GB2.8 GB
  • 4-bitweights 1.2 GB1.8 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.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · BF16 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

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.

    # Lester1996/Qwen3.5-2B-NoRA
    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 Lester1996/Qwen3.5-2B-NoRA --local-dir ~/models/Qwen3.5-2B-NoRA
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to Qwen3.5-2B-NoRA
  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. Qwen3.5-2B-Basefamily root
  2. Qwen3.5-2B
  3. Qwen3.5-2B-NoRA

New posts about versions of Qwen3.5-2B-Base (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 10All on Hugging Face
Merges 2All on Hugging Face

About the model

Lester1996/Qwen3.5-2B-NoRA
Author
Lester1996
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

370

Ask AI about this model

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

Comments