TIR-Qwen3.5-9B-NonThinking

This is the corrected, uninterrupted full-parameter SFT reproduction of the Qwen3.5-9B v1 non-thinking TIR recipe. It was trained for one contiguous epoch from Qwen/Qwen3.5-9B at a strict 32,768-token maximum length.

The checkpoint is in standard Hugging Face Transformers format and can be served with recent vLLM or Transformers releases that support Qwen3.5.

Training data

The final filtered v1 non-thinking mixture contains 20,338 rows and 50,001 unique referenced images:

Component	Rows before filtering
AgenticData true-multiturn SFT	4,561
CodeDance SFT	13,828
Cotrain attachment	2,200

The original 20,589-row v1 mixture was filtered to exclude 251 extreme-context examples using auditable image-count, aspect-ratio, text-length, and estimated-token thresholds. The final training file contains no assistant <think> blocks.

• Final rows: 20,338
• Training JSONL SHA-256: 1b64f18262d8571ee6d9d04a7653f64b7aa20bc11f26e6c5d62e39ee23d872cd
• Full filtering report: training_data/filter_report.json

Training configuration

• Base model: Qwen/Qwen3.5-9B
• Full-parameter SFT with ms-swift/Megatron
• One uninterrupted epoch from the base model
• Learning rate: 1e-5
• Minimum learning rate: 1e-6
• Warmup fraction: 0.05
• Maximum length: 32,768
• Global batch size: 64
• Micro batch size: 1
• Tensor parallel size: 2
• Pipeline parallel size: 1
• Vision tower and aligner frozen; language model trainable
• Non-thinking prefix enabled
• Weight dtype: bfloat16
• Final optimizer step/checkpoint: 317
• Training time: 8h 40m

The exported args.json records the complete training arguments.

Evaluation

A full 1,215-example TIR-Bench evaluation is running with four vLLM engines in agent mode, a Python sandbox, no task-specific prompts, rule-based scoring, a 32,768-token context window, up to 12,768 generated tokens, and thinking disabled. Final results are not claimed until that run and its error-only retry passes complete.

See evaluation/status.json for the evaluation protocol and status at publication time.

Inference

With HF_TOKEN set for this private repository, start a recent vLLM server:

vllm serve HEART77/TIR-Qwen3.5-9B-NonThinking \
  --port 8000 \
  --max-model-len 32768 \
  --reasoning-parser qwen3

Then run:

pip install openai
python inference.py --image /path/to/image.png --prompt "Describe the image and solve the problem."

inference.py also accepts text-only prompts when --image is omitted and explicitly disables thinking through the chat template.

Limitations

This checkpoint is specialized for multimodal tool-integrated reasoning. General-purpose safety, factuality, multilingual quality, and out-of-domain robustness have not been comprehensively evaluated. Generated code and tool calls must only be executed in a sandboxed environment.