Youtu-FT-V1-4B

Youtu-FT-V1-4B is a fine-tuned conversational model based on the Qwen 3.5 architecture. It is part of the Youtu-FT-V1 series, a family of reasoning-oriented models spanning 4B to 35B parameters, all distributed in GGUF format for efficient local inference.

Model Series

Key Features

• Base Architecture: Qwen 3.5 (qwen35)
• Context Length: 262,144 tokens
• Embedding Dimension: 2560
• Capabilities: Chat, tool use, reasoning with explicit thinking chains
• Sampling: temperature=0.1, top_p=0.9 (conservative, suited for precise tasks)
• Format: GGUF Q4_K_M quantization, compatible with llama.cpp, Ollama, LM Studio, and more
• License: MIT

Benchmarks

All benchmarks were run with Ollama using /api/generate raw mode (bypassing thinking tokens) on the same 50-sample subsets with identical few-shot prompts. This provides a fair comparison across model sizes without the confounding effect of reasoning-chain generation.

Observations

• ARC-Challenge (science reasoning): All three sizes achieve near-ceiling performance (94–96%), demonstrating strong structured reasoning across the series.
• PIQA (physical commonsense): Scales cleanly with model size (58% → 82% → 92%), with the MoE architecture showing a particularly strong gain.
• HellaSwag (commonsense completion): The 4B model unexpectedly outperforms the larger variants, likely due to better alignment between dense architecture and the raw-mode completion format used in this evaluation.
• WinoGrande (pronoun disambiguation): Performance peaks at the 9B dense model (66%). The MoE architecture does not show a benefit on this task under current settings.
• Inference speed: The 35B-A3B MoE model completes all 200 samples in 300 seconds — faster than the 9B dense model (498s) — benefiting from only 3B active parameters per token.

Note on thinking mode: All models in this series generate reasoning chains by default (thinking tokens). For benchmarks we used raw mode (raw: true) to obtain clean completions. In chat applications, you may observe additional <think> blocks that contain the model's reasoning process. This can be configured via the Ollama raw option or the --no-think flag in llama.cpp.

Quick Start

Ollama

ollama run hf.co/gatilin/Youtu-FT-V1-4B

To disable thinking tokens:

curl http://localhost:11434/api/generate -d '{
  "model": "hf.co/gatilin/Youtu-FT-V1-4B",
  "prompt": "你好",
  "raw": true
}'

llama.cpp

# Install
curl -LsSf https://llama.app/install.sh | sh

# Run with the model
llama serve -hf gatilin/Youtu-FT-V1-4B
llama cli -hf gatilin/Youtu-FT-V1-4B

llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
    repo_id="gatilin/Youtu-FT-V1-4B",
    filename="youtu-ft-v1-4b.gguf",
)

output = llm.create_chat_completion(
    messages=[{"role": "user", "content": "你好，请介绍一下自己"}]
)

Docker

docker model run hf.co/gatilin/Youtu-FT-V1-4B

Files

Citation

@misc{gatilin2025youtuft,
  title={Youtu-FT-V1: Fine-Tuned Reasoning Models},
  author={gatilin},
  year={2026},
  url={https://huggingface.co/gatilin/Youtu-FT-V1-4B}
}