Introducing VetJarvis 1.1-4B

언어 / Language: 🇰🇷 한국어 · 🇺🇸 English

We are pleased to release VetJarvis 1.1, an upgraded version featuring enhanced reasoning capabilities powered by an improved post-training process over the 1.0 release.

Quickstart

Serving VetJarvis-1.1-4B-Instruct

VetJarvis-1.1-4B-Instruct is a text-only LLM based on Qwen3.5-4B, and can be served in BF16 using a variety of frameworks including Hugging Face Transformers and vLLM.

Hugging Face Transformers

Qwen3.5 architecture is natively supported from transformers>=5.5.

pip install -U "transformers>=5.5" accelerate torch flash-attn

import torch
from transformers import AutoTokenizer, Qwen3_5ForConditionalGeneration

MODEL = "choonok/VetJarvis-1.1-4B-Instruct"

tokenizer = AutoTokenizer.from_pretrained(MODEL)
model = Qwen3_5ForConditionalGeneration.from_pretrained(
    MODEL,
    dtype=torch.bfloat16,
    device_map="auto",
    attn_implementation="flash_attention_2",
)
model.eval()

messages = [
    {
        "role": "system",
        "content": "You are 'VetJarvis', a clinical-support AI assistant for veterinarians.",
    },
    {"role": "user", "content": "Please tell me the metronidazole protocol."},
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True,
)

inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=32768,
        temperature=0.8,
        top_p=0.9,
        do_sample=True,
        pad_token_id=tokenizer.eos_token_id,
    )

print(tokenizer.decode(
    outputs[0][inputs.input_ids.shape[-1]:],
    skip_special_tokens=True,
))

On Blackwell GPUs (RTX 5090 / B200), a compatible build of flash-attn ≥ 2.7 is required. If you encounter build or compatibility issues, fall back to attn_implementation="sdpa".

vLLM

Recommended for production or high-throughput use cases. vllm>=0.18 supports the Qwen3.5 architecture, and the jointly trained MTP layer can be used for speculative decoding to improve throughput.

pip install "vllm>=0.18"

Launch an OpenAI-compatible server:

vllm serve choonok/VetJarvis-1.1-4B-Instruct \
    --served-model-name VetJarvis-1.1-4B-Instruct \
    --port 8000 \
    --max-model-len 65535 \
    --dtype bfloat16 \
    --gpu-memory-utilization 0.85 \
    --speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":2}'

Example:

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:8000/v1",
    api_key="EMPTY",
)

response = client.chat.completions.create(
    model="VetJarvis-1.1-4B-Instruct",
    messages=[
        {
            "role": "system",
            "content": "You are 'VetJarvis', a clinical-support AI assistant for veterinarians.",
        },
        {
            "role": "user",
            "content": "Please tell me the metronidazole protocol.",
        },
    ],
    max_tokens=32768,
    temperature=0.8,
    top_p=0.9,
    extra_body={
        "chat_template_kwargs": {"enable_thinking": True},
    },
)

print(response.choices[0].message.content)

Recommended Parameters

Key Improvements

• Enhanced Reasoning Performance: While VetJarvis 1.0 showed minimal performance differences with or without the think option enabled, version 1.1 addresses this limitation and delivers meaningful performance gains in think mode.
• GPT-5.4-mini Class Reasoning: Through a refined post-training pipeline, VetJarvis 1.1 provides logical reasoning performance on par with GPT-5.4-mini.

Usage

For usage instructions and detailed model documentation, please refer to the VetJarvis 1.0 repository.