robotflowlabs/qwen2.5-7b-instruct-int4

Qwen2.5-7B-Instruct — INT4 NF4 Quantized

Alibaba's Qwen2.5-7B-Instruct quantized to 4-bit NF4 with double quantization for robotic reasoning and planning. 2.7x smaller — from 14.5 GB to 5.3 GB — while preserving instruction-following and reasoning capabilities.

This model is part of the RobotFlowLabs model library, built for the ANIMA agentic robotics platform — a modular ROS2-native AI system that brings foundation model intelligence to real robots operating in the real world.

Why This Model Exists

Robots need to reason about instructions, plan multi-step tasks, and generate structured outputs — all in real-time on edge hardware. Qwen2.5-7B is one of the strongest open-source instruction-following models at this scale, with excellent performance on reasoning, coding, and structured output generation. At 14.5 GB it's too large for edge GPUs. INT4 NF4 double quantization brings it to 5.3 GB — fitting on a single L4 24GB alongside vision models.

Model Details

Property	Value
Architecture	Qwen2 (decoder-only transformer)
Parameters	7B
Hidden Dimension	3584
Layers	28
Attention Heads	28 (4 KV heads, GQA)
MLP Dimension	18944 (SiLU activation)
Context Length	32,768 tokens
Vocabulary	152,064 tokens
RoPE	θ = 1,000,000
Quantization	NF4 double quantization (bitsandbytes)
Original Model	Qwen/Qwen2.5-7B-Instruct
License	Apache-2.0

Compression Results

Quantized on an NVIDIA L4 24GB GPU using bitsandbytes NF4 with double quantization.

Metric	Original	INT4 Quantized	Change
Total Size	14,537 MB	5,301 MB	2.7x smaller
Quantization	BF16	NF4 + double quant	4-bit weights
Compute Dtype	BF16	BF16	Preserved at inference
Format	SafeTensors	SafeTensors	Direct HF loading

Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "robotflowlabs/qwen2.5-7b-instruct-int4",
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("robotflowlabs/qwen2.5-7b-instruct-int4")

messages = [
    {"role": "system", "content": "You are a robotic task planner."},
    {"role": "user", "content": "Plan the steps to pick up the red cup and place it on the shelf."}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

With FORGE (ANIMA Integration)

from forge.language import LanguageModelRegistry

planner = LanguageModelRegistry.load("qwen2.5-7b-instruct-int4")
plan = planner.generate("Pick up the red cup and place it on the shelf")

Use Cases in ANIMA

Qwen2.5-7B serves as the reasoning backbone in ANIMA:

• Task Planning — Decompose natural language instructions into executable step sequences
• Code Generation — Generate robot control scripts and action sequences
• Structured Output — Produce JSON task plans, waypoint lists, and parameter configs
• Safety Reasoning — Evaluate whether proposed actions are safe before execution
• Error Recovery — Diagnose failures and generate recovery plans
• Human Dialogue — Natural language interaction with operators

About ANIMA

ANIMA is a modular, ROS2-native agentic robotics platform developed by RobotFlowLabs. It combines 58 specialized AI modules into a unified system for real-world robotic autonomy.

Other Collections

• ANIMA Vision — SAM2, DINOv2, CLIP, SigLIP, Depth Anything
• ANIMA Language — Qwen2.5, SmolLM2
• ANIMA VLM — Qwen2.5-VL
• ANIMA VLA — SmolVLA, RDT2-FM, FORGE students

Intended Use

Designed For

• On-device robotic task planning and reasoning
• Instruction following in manipulation and navigation pipelines
• Structured output generation (JSON, code, action sequences)
• Multi-turn dialogue with human operators

Limitations

• INT4 quantization may slightly reduce performance on complex reasoning benchmarks
• 32K context window may not be sufficient for very long interaction histories
• Requires GPU (bitsandbytes NF4 does not run on CPU)
• Inherits biases from Qwen2.5 training data

Out of Scope

• Safety-critical autonomous decision making without human oversight
• Medical or legal advice generation
• Generation of harmful content

Technical Details

Compression Pipeline

Original Qwen2.5-7B-Instruct (BF16, 14.5 GB)
    │
    └─→ bitsandbytes NF4 double quantization
        ├─→ bnb_4bit_quant_type: nf4
        ├─→ bnb_4bit_use_double_quant: true
        ├─→ bnb_4bit_compute_dtype: bfloat16
        └─→ model.safetensors (5.3 GB)

• Quantization: NF4 (Normal Float 4-bit) with double quantization via bitsandbytes
• Compute: BF16 at inference — weights dequantized on-the-fly
• Hardware: NVIDIA L4 24GB, CUDA 13.0, PyTorch 2.10, Python 3.14

Attribution

• Original Model: Qwen/Qwen2.5-7B-Instruct by Alibaba Cloud
• License: Apache-2.0
• Paper: Qwen2.5 Technical Report — Qwen Team, 2024
• Compressed by: RobotFlowLabs using FORGE

Citation

@article{qwen2.5,
  title={Qwen2.5 Technical Report},
  author={Qwen Team},
  journal={arXiv preprint arXiv:2412.15115},
  year={2024}
}

---

Built with FORGE by RobotFlowLabs
Optimizing foundation models for real robots.