You can deploy this model and start earning money today!

gatilin/
Youtu-FT-V1-9B

Name: Youtu-FT-V1-9B
Availability: InStock

Views

No views yet

Youtu-FT-V1-9B

Youtu-FT-V1-9B is a fine-tuned conversational model based on the Qwen 3.5 architecture. It is the mid-sized member 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

Model	Parameters	Quantization	Size	Architecture
Youtu-FT-V1-4B	4.2B	Q4_K_M	2.7 GB	Dense
Youtu-FT-V1-9B	9.0B	Q4_K_M	5.6 GB	Dense
Youtu-FT-V1-35B-A3B	35B (3B active)	Q4_K_M	21 GB	MoE

Key Features

Base Architecture: Qwen 3.5 (qwen35)
Parameters: 9.0B
Context Length: 262,144 tokens
Embedding Dimension: 4096
Capabilities: Chat, tool use, reasoning with explicit thinking chains
Sampling: temperature=0.7, top_p=0.9
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.

Benchmark	Shot	Youtu-FT-V1-4B	Youtu-FT-V1-9B	Youtu-FT-V1-35B-A3B
ARC-Challenge	3-shot	94.0%	96.0%	96.0%
HellaSwag	2-shot	62.0%	42.0%	44.0%
PIQA	1-shot	58.0%	82.0%	92.0%
WinoGrande	0-shot	58.0%	66.0%	52.0%
Average	—	68.0%	71.5%	71.0%

Observations

ARC-Challenge (science reasoning): Near-perfect at 96%, matching the 35B MoE variant. A +2pp improvement over the 4B model.
PIQA (physical commonsense): Jump to 82% from 58% at 4B — the largest single-model improvement in the series, demonstrating the 9B dense architecture's sweet spot for physical reasoning.
WinoGrande (pronoun disambiguation): Peaks at 66% — the best result across all three model sizes, indicating dense 9B is optimal for this task.
HellaSwag (commonsense completion): 42%, notably lower than the 4B (62%). This pattern is shared by the 35B variant and likely reflects a tension between the thinking-chain training objective and pure completion tasks under raw mode evaluation.
Inference speed: 498 seconds for 200 samples — significantly slower than the 35B MoE (300s) due to all 9B parameters being active on every 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-V1-9B

To disable thinking tokens:

bash

curl http://localhost:11434/api/generate -d '{
  "model": "hf.co/gatilin/Youtu-FT-V1-9B",
  "prompt": "你好",
  "raw": true
}'

llama.cpp

bash

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

# Run with the model
llama serve -hf gatilin/Youtu-FT-V1-9B
llama cli -hf gatilin/Youtu-FT-V1-9B

llama-cpp-python

python

from llama_cpp import Llama

llm = Llama.from_pretrained(
    repo_id="gatilin/Youtu-FT-V1-9B",
    filename="youtu-ft-v1-9b-Q4_K_M.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-V1-9B

Files

File	Description
`youtu-ft-v1-9b-Q4_K_M.gguf`	GGUF model weights (Q4_K_M, 5.6 GB)
`config.json`	Model format and family metadata
`params.json`	Default sampling parameters (temperature=0.7, top_p=0.9)
`template.txt`	Chat template for the model
`Modelfile`	Ollama Modelfile for local use
`manifest.json`	Ollama manifest for verification
`LICENSE`	MIT license

Citation

bibtex

@misc{gatilin2025youtuft,
  title={Youtu-FT-V1: Fine-Tuned Reasoning Models},
  author={gatilin},
  year={2026},
  url={https://huggingface.co/gatilin/Youtu-FT-V1-9B}
}