--- title: Taimi-14B-Med canonical_url: "https://www.modelscope.ai/models/TMiRob/Taimi-14B-Med" md_url: "https://www.modelscope.ai/models/TMiRob/Taimi-14B-Med.md" repository: TMiRob/Taimi-14B-Med last_updated: 2026-08-20 license: apache-2.0 pipeline_tag: text-generation tasks: - text-generation model_type: - qwen2 architectures: - Qwen2ForCausalLM base_model: - Qwen/Qwen2.5-14B-Instruct-AWQ base_model_relation: quantized parameters: 3.3B tensor_type: - F16 - I32 library_name: - transformer - safetensors - pytorch frameworks: - pytorch language: - en downloads: 63 stars: 0 tags: - chat - medical --- # Taimi-14B-Med > Taimi-14B-Med - An open-source model by TMiRob on ModelScope. Taimi-14B-Med is a medical-specialized large language model developed by Taimi, planned as a post-trained (fine-tuned) derivative of Qwen2.5-14B-Instruct-AWQ. It is designed for deployment in… TMiRob/Taimi-14B-Med is a 3.3B-parameter text-generation model on ModelScope. licensed under apache-2.0. derived from Qwen/Qwen2.5-14B-Instruct-AWQ. - **Repository**: TMiRob/Taimi-14B-Med - **License**: apache-2.0 - **Tasks**: text-generation - **Parameters**: 3.3B - **Base model**: Qwen/Qwen2.5-14B-Instruct-AWQ - **Tags**: chat, medical - **Downloads**: 63 - **Stars**: 0 - **Last updated**: 2026-08-20 Source: https://www.modelscope.ai/models/TMiRob/Taimi-14B-Med --- # Taimi-14B-Med **Taimi-14B-Med** is a **medical-specialized large language model** developed by Taimi, planned as a post-trained (fine-tuned) derivative of Qwen2.5-14B-Instruct-AWQ. It is designed for deployment in healthcare institutions and on service robots: medical indicator Q&A, nursing guidance, abnormal indicator alerting, and automated nursing record registration. > **Version note**: Current release **v0.1.0** consists of the base-model weights with a rebranded configuration. Post-trained medical weights will be released in a later version (v0.2.0+), and this card will be updated with fresh evaluation results at that time. ## Introduction Taimi-14B-Med serves as the **text brain** for medical scenarios: service robots handle front-end scene capture, visual acquisition, and physical data acquisition, combined with dedicated visual processing tools (e.g., YOLO) to output structured medical data; Taimi-14B-Med then applies its medical knowledge to answer patient-specific medical indicator questions, provide nursing guidance, raise abnormal indicator alerts, and automate nursing record registration. The model has been evaluated on publicly available Chinese medical benchmarks — CMB (11,200 questions), CMExam (6,811 questions), and CMDD (human-reviewed sampled dialogues) — demonstrating solid Chinese medical Q&A performance at the 14B scale (see Evaluation & Performance). ## Model information | Item | Value | |---|---| | Base model | `Qwen/Qwen2.5-14B-Instruct-AWQ` (Apache 2.0) | | Parameters | 14.7B (13.1B non-embedding) | | Quantization | AWQ 4-bit (10.31GB weights) | | Architecture | Causal LM, transformers (RoPE / SwiGLU / RMSNorm / GQA: 40 Q heads · 8 KV heads, 48 layers) | | Context length | Up to 32768 by default (deployed at 4096 in this project) | | Serving framework | vLLM (OpenAI-compatible API, `--served-model-name taimi-14b-med`) | | License | Apache 2.0 (inherited from base model) | ## Requirements - **Serving (recommended)**: vLLM (official Docker image `vllm/vllm-openai:latest` bundles the CUDA runtime; no host CUDA installation needed); NVIDIA driver + nvidia-container-toolkit. - **Transformers direct loading**: transformers >= 4.37.0. - **Hardware**: GPU VRAM >= 16GB (24GB recommended, covering 32K context and higher concurrency); system RAM >= 32GB; disk >= 20GB free (10.31GB weights). ## Quickstart ### Option 1: Local serving (vLLM, recommended) Download the weights from ModelScope (`taimi/Taimi-14B-Med`), then point `-v` to your local weights directory: ```bash docker run -d --name taimi-14b-med --gpus all --entrypoint vllm \ -v /path/to/Taimi-14B-Med:/models/taimi-14b-med \ -p 8002:8002 vllm/vllm-openai:latest serve /models/taimi-14b-med \ --port 8002 --gpu-memory-utilization 0.85 --max-model-len 4096 \ --kv-cache-dtype fp8 --max-num-seqs 16 --served-model-name taimi-14b-med ``` To download from ModelScope: ```bash pip install modelscope modelscope download --model taimi/Taimi-14B-Med --local_dir /path/to/Taimi-14B-Med ``` ### Option 2: API calls (OpenAI-compatible) ```bash curl http://localhost:8002/v1/chat/completions \ -H "Content-Type: application/json" \ -d '{"model":"taimi-14b-med","messages":[{"role":"user","content":"After radionuclide therapy, what contact restrictions apply to family members when the patient is discharged?"}],"max_tokens":200}' ``` ### Option 3: Direct loading with Transformers ```python from transformers import AutoModelForCausalLM, AutoTokenizer model_name = "/path/to/Taimi-14B-Med" # local directory after downloading from ModelScope model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto", device_map="auto") tokenizer = AutoTokenizer.from_pretrained(model_name) messages = [{"role": "user", "content": "Explain the meaning of the 25 uSv/h threshold in nuclear medicine discharge assessment."}] text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) inputs = tokenizer([text], return_tensors="pt").to(model.device) outputs = model.generate(**inputs, max_new_tokens=200) print(tokenizer.decode(outputs[0], skip_special_tokens=True)) ``` ## Evaluation & Performance Evaluated on **publicly available Chinese medical benchmarks** (temperature=0): | Benchmark | Scale | Metric | Taimi-14B-Med (v0.1.0) | Reference | |---|---|---|---|---| | CMB-test (Chinese Medical Benchmark) | 11,200 questions | ACC | **74.0%** (single-choice 77.1% / multi-choice 47.2%) | Qwen-72B-Chat 74.38% (community) | | CMExam (Chinese physician licensing exam) | 6,811 questions | ACC | **77.4%** | Paper baselines: GPT-4 61.6%, human 71.6% | | CMDD (Chinese medical dialogue dataset) | 100 sampled dialogues | Human review (relevance / safety / refusal appropriateness, 5-point scale, double-scored) | Overall ~60, stable across cases | Physician reference answers ~70 | Key notes: - CMB covers 6 sub-domains (nursing 79.1%, pharmacy 76.1%, physician 72.9%, postgraduate 71.7%, professional knowledge 70.9%, medical technology 70.8%). - CMExam exceeds both paper baselines (GPT-4 61.6%, human 71.6%); top departments include tuberculosis 90.0%, psychiatry 86.4%, oncology 85.9%. - CMB / CMExam were released in 2023 and may overlap with the base model's training corpus; ACC results serve as reproducibility validation (the 14B model matches the community-reported 72B-Chat level on CMB). - CMDD human review indicates high stability (low variance across cases), suitable for pre-diagnosis screening; physician reference answers score higher on average but with larger variance. **Resource usage** (measured 2026-08-14, RTX 5080 16GB): - VRAM 12537MiB / 16303MiB (77%, `--gpu-memory-utilization 0.85` + fp8 KV cache) - System RAM ~6.1GB; model load ~45-90s - Context 4096 (deployment config); `/v1/models` reports `taimi-14b-med` after serving **Note**: v0.1.0 contains base-model weights; the evaluation above reflects base-model capability (Chinese medical Q&A performance). After post-training (v0.2.0+), the table will be re-evaluated and updated. ## Citation If you use this model, please cite: ```text Taimi. Taimi-14B-Med: A Medical-Specialized Large Language Model (derived from Qwen2.5-14B-Instruct-AWQ). 2026. ``` ## License This model is a derivative work post-trained (fine-tuned) from Qwen2.5-14B-Instruct-AWQ (Apache 2.0), released under the **Apache 2.0** license. The LICENSE file is retained in the repository; please comply with the base-model license terms when using it.