Instructions to use Glasseek/Glasseeker-9B-SFT with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Glasseek/Glasseeker-9B-SFT with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="Glasseek/Glasseeker-9B-SFT")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("Glasseek/Glasseeker-9B-SFT")
model = AutoModelForMultimodalLM.from_pretrained("Glasseek/Glasseeker-9B-SFT", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Glasseek/Glasseeker-9B-SFT with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Glasseek/Glasseeker-9B-SFT"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Glasseek/Glasseeker-9B-SFT",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Glasseek/Glasseeker-9B-SFT

SGLang

How to use Glasseek/Glasseeker-9B-SFT with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Glasseek/Glasseeker-9B-SFT" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Glasseek/Glasseeker-9B-SFT",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Glasseek/Glasseeker-9B-SFT" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Glasseek/Glasseeker-9B-SFT",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Glasseek/Glasseeker-9B-SFT with Docker Model Runner:
```
docker model run hf.co/Glasseek/Glasseeker-9B-SFT
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

Glasseeker-9B-SFT

A Transparent Recipe for Training and Trustworthy Evaluation of DeepSearch Agents

Project Repository | SFT Dataset

Glasseeker-9B-SFT is the 9B supervised fine-tuning checkpoint of our deep search agent, built on Qwen3.5-9B. It learns multi-step search planning, query rewriting, evidence retrieval, and answer synthesis from search trajectories.

Glasseek connects verifiable data construction, reinforcement learning with context management, and evaluation with anti-leakage safeguards. These checkpoints initialize the reinforcement-learning stage described in Glasseek: A Transparent Recipe for Training and Trustworthy Evaluation of DeepSearch Agents.

Highlights

Compact search agents. Two model sizes support research on deep search at the 4B and 9B scales.
Long-context SFT. Sequence packing and a maximum training sequence length of 131,072 tokens support supervision over extended interactions.
Shared research resources. The companion SFT dataset contains 5,762 trajectories covering 4,448 unique questions.

Model Family

Checkpoint	Base model	Training stage	Weight format
Glasseeker-4B-SFT	Qwen3.5-4B	Supervised fine-tuning	Safetensors
Glasseeker-9B-SFT	Qwen3.5-9B	Supervised fine-tuning	Safetensors

Each model repository contains model weights, configuration files, the tokenizer, and the chat template at its root.

SFT Benchmark Results

The following comparison combines the paper's 9B SFT baseline in Table 3 with the same-scale baselines in Table 2. Scores are accuracy (%).

Model	BrowseComp	BrowseComp-ZH	Seal-0	XDS-2510	GAIA-Text
Glasseek-9B-SFT	55.5	59.5	34.2	44.0	63.1
Apodex-1.0-4B-SFT	48.8	63.5	--	--	--
QUEST-9B-SFT	45.4	--	--	--	78.6
QUEST-4B-SFT	40.0	--	--	--	77.7
Marco-DeepResearch-8B	31.4	47.1	--	42.0	69.9
DeepSearch-World-9B	31.2	36.4	--	49.0	61.5
MiroThinker-v1.0-8B	31.1	40.2	40.4	34.0	66.4
Qwen3.5-9B	18.3	32.1	28.8	36.0	60.2

Glasseek-9B-SFT is evaluated with Hermes-agent, the paper's web-search and webpage-retrieval tools, and anti-leakage safeguards. Its BrowseComp score uses the fixed 200-question ablation subset, averaged over three independent runs. The other benchmarks use their full evaluation sets; GAIA-Text is the 103-question text-only subset. Qwen3.5-9B scores are obtained with our harness. Other baseline scores follow their original reports and evaluation settings. -- and n/r denote unreported results. Baselines retain their reported training stages.

The underlying values are available in CSV format.

Training Configuration

The paper reports the following SFT configuration for the 9B model in Appendix C:

Setting	Value
Training budget	3 epochs
Learning rate	2e-5
Minimum learning rate	2e-6
Warmup ratio	0.03
Global batch size	64 packed sequences
Maximum sequence length	131,072 tokens
Sequence packing	Enabled
Training resources	16 GPUs, approximately 15 hours

From SFT to CM-GSPO

The Glasseek pipeline continues from SFT with Compaction-Managed GSPO (CM-GSPO). Context compaction divides a rollout into execution sub-trajectories with recorded conditioning histories. CM-GSPO uses length-normalized sub-trajectory likelihood ratios and length-based advantage reweighting to balance contributions within each original rollout. Summary-generation tokens are excluded from the RL objective.

Quick Start

Use an inference runtime with Qwen3.5 support. For example, serve this checkpoint directly from Hugging Face with vLLM:

vllm serve Glasseek/Glasseeker-9B-SFT \
    --served-model-name glasseek-sft-9b \
    --max-model-len 131072 \
    --reasoning-parser qwen3 \
    --enable-auto-tool-choice \
    --tool-call-parser qwen3_coder

For the 4B checkpoint, use Glasseek/Glasseeker-4B-SFT and the served model name glasseek-sft-4b.

For multi-step search, connect the model endpoint to the Hermes evaluation backend in the project repository. The harness executes tool calls and returns observations to the model. The paper uses Serper for web search and Jina Reader for webpage retrieval.

Related Resources

Project repository: project overview, evaluation runner, and Hermes backend.
Glasseek training data: 5,762 SFT trajectories covering 4,448 unique questions.

Downloads last month: 62

Safetensors

Model size

9B params

Tensor type

BF16

Model tree for Glasseek/Glasseeker-9B-SFT

Quantizations

1 model