Verus-r1

This repository contains model weights and configuration files for Verus-r1 in the Hugging Face Transformers format.

Compatible with Hugging Face Transformers, vLLM, SGLang, and other major inference frameworks.

Built for coding, reasoning, debugging, and concise general assistance.

Verus-r1 Highlights

• Coding-Focused: Writes, fixes, explains, and reviews code.
• Reasoning-Oriented: Works through multi-step problems clearly.
• Long Context: Can handle large prompts, files, and long conversations.
• Instruction Following: Responds in the format and style requested.
• Efficient: A compact 2B model for local or hosted inference.

Model Overview

Property	Value
Parameters	~2B
Context Length	262,144 tokens
Architecture	Qwen3.5
Chat Format	ChatML (<|im_start|> / <|im_end|>)
Dtype	bfloat16
License	Apache 2.0

Quickstart

Installation

pip install "transformers>=4.52.0" accelerate torch

Code Generation

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

MODEL_ID = "8F-ai/Verus-r1"

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
model.eval()

messages = [
    {
        "role": "system",
        "content": "You are Verus-r1, a reasoning coding assistant made by 8F-ai. You think through problems carefully before responding."
    },
    {
        "role": "user",
        "content": "Write a Python async context manager that manages a PostgreSQL connection pool using asyncpg."
    }
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.inference_mode():
    generated_ids = model.generate(**inputs, max_new_tokens=2048, temperature=0.6, top_p=0.95)

output = tokenizer.decode(generated_ids[0][len(inputs.input_ids[0]):], skip_special_tokens=True)
print(output)

Quantized Inference (4-bit NF4, ~2 GB VRAM)

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
import torch

quantization_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4",
)

tokenizer = AutoTokenizer.from_pretrained("8F-ai/Verus-r1")
model = AutoModelForCausalLM.from_pretrained(
    "8F-ai/Verus-r1",
    quantization_config=quantization_config,
    device_map="auto",
)

Intended Use Cases

Use Case	Example
Code Generation	Write functions, classes, and scripts
Debugging	Fix bugs from code or error messages
Code Review	Explain code and suggest improvements
Reasoning	Break down multi-step problems
Long Context	Work with long prompts and files
General Q&A	Answer clearly and concisely

Limitations

• English-Primary: Fine-tuning was conducted predominantly on English-language code and documentation.

Citation

@misc{verusr12026,
  title        = {Verus-r1: A Reasoning-Focused Coding Language Model with 262K Context},
  author       = {8F-ai},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/8F-ai/Verus-r1}},
  note         = {Apache 2.0 License}
}

License

Verus-r1 is released under the Apache License 2.0. See LICENSE for full terms.

Derived from Qwen/Qwen3.5-2B (Apache 2.0).

---

_{Built by the 8F-ai Team}