HyWe-35B

HyWe-35B is the fine-tuned agent from HyWe: Unifying Search and Operation for Real-World Hybrid Web Tasks. It combines text-based search and rendered browser operation in one tool-using agent, using switch to change modes. The checkpoint is fine-tuned from Qwen3.6-35B-A3B and is stored in BF16 across 16 safetensors shards (approximately 70.2 GB including configuration files).

💻 Code, setup, and evaluation · 🤗 HyWeBench

Use with the HyWe framework

vllm serve yzxjb/HyWe-35B --served-model-name HyWe-35B \
  --host 127.0.0.1 --port 8000 --tensor-parallel-size 2 --max-model-len 224000

Use a vLLM build supporting the checkpoint's qwen3_5_moe architecture and adjust GPU count/concurrency to available memory. Set MODEL=HyWe-35B and BASE_URL=http://127.0.0.1:8000/v1 in the code repository's example config, then run scripts/run_hywe.sh. The framework supplies tool schemas, executes calls, switches the S/O prompt, and manages context folding.

Model outputs use <think>...</think> followed by <tool_call>{"name": "...", "arguments": {...}}</tool_call> or <answer>...</answer>. Browser screenshots have a 1280×720 viewport; coordinates use a 0–1000 grid. A standard chat UI without the HyWe tools and context renderer does not reproduce the benchmark setup.

Data and training

• 🤗 HyWe-Hybrid-SFT-Data: 3,735 hybrid trajectories, 8,055 rendered segments and 45,045 original screenshots.
• 🤗 HyWe-Search-SFT-Data: the 128K multi-turn conversion of 7,000 XYZ-Aquila-SFT trajectories.

Training used full language-model fine-tuning with the vision encoder and projector frozen, a 131,072-token sequence length, BF16, an effective batch size of 32, peak learning rate 2e-5, and one configured epoch. Portable saved settings and their provenance limits are documented in training/ in the code release. Exact GPU type, elapsed training time and source sampling proportions were not recorded in the saved checkpoint arguments.

Evaluation

HyWeBench metric	Result (%)
All checklist items achieved	47.0
Information items achieved	82.2
Operation items achieved	68.8

These results aggregate archived judgments for the 100-task benchmark, using the released framework and Kimi-K3 checklist judge. The code repository includes all 14 comparison/ablation verdict sets and a script to recompute their rates. New live-site runs can differ because websites, search results and model services change. The recorded runs used Jina Search; the paper's appendix describes Serper, a discrepancy documented in the release.

Limitations

The agent may produce incorrect answers or actions and can fail on CAPTCHA, authentication, changing pages, and long-horizon tasks. Local scripted smoke tests validate integration, not learned task-solving performance. Use the release's benchmark protocols when comparing models.

The checkpoint is distributed under Apache-2.0, consistent with its base model.