Youtu-FT-V1-35B-A3B

Youtu-FT-V1-35B-A3B is a fine-tuned Mixture-of-Experts (MoE) conversational model based on the Qwen 3.5 MoE architecture. It is the largest member of the Youtu-FT-V1 series, with 34.7B total parameters but only ~3B active per token, delivering a compelling balance of reasoning quality and inference speed.

Model Series

Key Features

• Base Architecture: Qwen 3.5 MoE (qwen35moe)
• Parameters: 34.7B total, ~3B active per token (sparse activation)
• Context Length: 262,144 tokens
• Embedding Dimension: 2048
• Capabilities: Chat, tool use, reasoning with explicit thinking chains
• Sampling: temperature=1.0, top_p=0.95, top_k=20, presence_penalty=1.5
• 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): Matches the 9B at 96%, demonstrating that the additional expert capacity does not compromise structured reasoning tasks.
• PIQA (physical commonsense): The standout result at 92% — a +10pp gain over the 9B, and the largest single-benchmark score in the series. The MoE architecture with diverse expert knowledge appears to particularly benefit physical commonsense tasks.
• HellaSwag (commonsense completion): 44%, consistent with the 9B (42%). The MoE does not overcome the tension between reasoning-chain training and raw-mode completion evaluation.
• WinoGrande (pronoun disambiguation): 52%, the lowest in the series. Sparse activation may dilute the context-tracking needed for precise pronoun resolution.
• Inference speed: 300 seconds for 200 samples — the fastest in the series, 40% quicker than the 9B (498s). This is the MoE advantage: only ~3B parameters are active per forward pass, making it faster than the 9B dense model while carrying far more stored knowledge.

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-V1-35B-A3B

To disable thinking tokens:

curl http://localhost:11434/api/generate -d '{
  "model": "hf.co/gatilin/Youtu-FT-V1-35B-A3B",
  "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-V1-35B-A3B
llama cli -hf gatilin/Youtu-FT-V1-35B-A3B

llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
    repo_id="gatilin/Youtu-FT-V1-35B-A3B",
    filename="Youtu-FT-V1-35B-A3B.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-V1-35B-A3B

Files

Citation

@misc{gatilin2025youtuft,
  title={Youtu-FT-V1: Fine-Tuned Reasoning Models},
  author={gatilin},
  year={2026},
  url={https://huggingface.co/gatilin/Youtu-FT-V1-35B-A3B}
}