---
title: Agentic-TCM
canonical_url: "https://www.modelscope.ai/models/cehpe2-wodvej-jefxoP/Agentic-TCM"
md_url: "https://www.modelscope.ai/models/cehpe2-wodvej-jefxoP/Agentic-TCM.md"
repository: cehpe2-wodvej-jefxoP/Agentic-TCM
last_updated: 2026-07-22
pipeline_tag: image-text-to-text
tasks:
  - image-text-to-text
model_type:
  - qwen3_5_moe
architectures:
  - Qwen3_5MoeForConditionalGeneration
parameters: 6.0B
tensor_type:
  - F16
  - I32
  - BF16
library_name:
  - safetensors
  - pytorch
frameworks:
  - Pytorch
downloads: 3
stars: 0
---

# Agentic-TCM

> Agentic-TCM - An open-source model by cehpe2-wodvej-jefxoP on ModelScope. Qwen3_5MoeForConditionalGeneration VLM+MoE，40层，256 Experts（Top-8），Hybrid Linear/Full Attention，262K Context，1152维视觉塔；基于15000条16来源混合数据SFT（TCM、Tool Calling、GUI…

- **Repository**: cehpe2-wodvej-jefxoP/Agentic-TCM
- **Tasks**: image-text-to-text
- **Parameters**: 6.0B
- **Downloads**: 3
- **Stars**: 0
- **Last updated**: 2026-07-22

Source: https://www.modelscope.ai/models/cehpe2-wodvej-jefxoP/Agentic-TCM

---

# Qwen3.5-MoE-TCM-VLM

## Overview

Qwen3.5-MoE-TCM-VLM is a domain-specialized Vision-Language Model based on Qwen3_5MoeForConditionalGeneration.

The model is designed for Traditional Chinese Medicine (TCM), visual understanding, tool calling, GUI Agent tasks and long-context reasoning.

---

## Base Model

- Architecture: Qwen3_5MoeForConditionalGeneration
- Type: Vision Language Model (VLM)
- Mixture of Experts (MoE)
- Layers: 40
- Experts: 256 routed experts
- Activated Experts: Top-8 per token
- Hybrid Attention:
  - Linear Attention
  - Full Attention (1 every 4 layers)
- Max Context Length: 262,144 tokens
- Vision Tower Hidden Size: 1152

---

## Training

### Dataset

15,000 samples mixed from 16 datasets.

Domain distribution:

| Domain | Samples |
|---------|---------:|
| TCM | 6,800 |
| Tool / Function Calling | 2,800 |
| General Dialogue | 3,250 |
| General Vision | 1,500 |
| GUI / Agent | 650 |

Main datasets include:

- tcm_vision_sft_35k
- tcm_ladder_10k
- tcm_text_sft_25k
- finevision_replay_23k
- sharegpt_multi_turn_zh
- hermes_func_call_mt
- apigen_mt_5k
- taskbench_tool_use
- xlam_func_call_60k
- showui_gui_agent
- mm_mind2web_agent

---

### Fine-tuning

LoRA configuration

- Rank: 64
- Alpha: 128

LoRA is applied to:

- MoE Experts
- Linear Attention
- Full Attention

=

Training loss:

- Start: 1.62
- End: 0.69

---

## Training Hardware

- 4 × NVIDIA DGX Spark
- Custom SFT framework optimized for DGX Spark

---

## Quantization

AutoRound 0.14.2

Configuration:

- W4A16
- Group Size: 128
- Symmetric Quantization

Quantized modules:

- language_model.layers

Kept in BF16 / FP:

- Vision Tower
- Router Gate
- LM Head

Calibration:

- 128 proportional calibration samples

Model size:

- ~20 GB
- 5 shards

---

## Deployment

Inference Engine:

- vLLM

Tensor Parallel:

- TP=2

Hardware:

- 2 × DGX Spark

Verified:

- 256K Context Inference

---

## Intended Use

Designed for:

- Traditional Chinese Medicine
- Vision-language reasoning
- Tool Calling
- GUI Agent
- Web Agent
- Long-context applications

---

## License

Please follow the license terms of the original Qwen3.5 model.