GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
lilliii
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
lilliii
Model Tree
Input Modalities
Output Modalities
Supported Functionality
FengHe is a meteorological service domain large language model developed by the Public Meteorological Service Center of China Meteorological Administration based on GLM-4.5-Air. It is further trained with over 50 million tokens of high-quality meteorological service corpus and 490,000 scenario-oriented instruction-tuning examples.
FengHe is the first large language model in China specifically designed for the meteorological service domain. It is optimized for meteorological service requirement understanding, service-oriented content generation, meteorological reasoning and decision support, and meteorological service tool calling.
This repository provides the complete FengHe model release, including the Model Card, model weights, configuration files, tokenizer files, training data description, and evaluation result visualizations.
The model can be loaded directly with Transformers, vLLM, or SGLang.
| Item | Description |
|---|---|
| Model name | FengHe |
| Model type | Meteorological service domain large language model |
| Base model | GLM-4.5-Air |
| Architecture | Mixture-of-Experts, Causal Language Model |
| Total parameters | 106B |
| Activated parameters | 12B |
| Context length | 128K tokens |
| Training precision | BF16 |
| License | MIT |
| Languages | Chinese, English |
| Domain | Meteorological services, weather forecasting, warning services, decision support, tool calling |
FengHe inherits the MoE architecture and hybrid reasoning capability of GLM. It supports reasoning-oriented generation for complex tasks and direct-response generation for real-time interactions.
| Parameter | Value |
|---|---|
| Architecture | Glm4MoeForCausalLM |
| Model type | glm4_moe |
| Number of hidden layers | 46 |
| Hidden size | 4096 |
| Intermediate size | 10944 |
| MoE intermediate size | 1408 |
| Attention heads | 96 |
| Key-value heads | 8 |
| Routed experts | 128 |
| Shared experts | 1 |
| Experts per token | 8 |
| Head dimension | 128 |
| Max position embeddings | 131072 |
| Vocabulary size | 151552 |
| Activation function | SiLU |
| Normalization | RMSNorm |
| RoPE theta | 1000000 |
| Torch dtype | BF16 |
FengHe is trained on a high-quality meteorological service corpus built from multiple professional data sources, including:
Based on these sources, we constructed a high-quality meteorological service corpus of over 50 million tokens and manually annotated 490,000 scenario-oriented meteorological instruction examples. These data improve the model's professional expression, business understanding, reasoning, decision-support capability, and tool-calling ability in the meteorological service domain.
FengHe is optimized for meteorological service scenarios, with improvements in the following capabilities:
Meteorological service requirement understanding Understands user needs in weather forecasting, severe weather, warning services, event support, public services, and decision-support scenarios.
Meteorological service content generation Generates service-oriented meteorological texts for the public, industry users, and decision makers, including weather briefings, service reports, risk alerts, and warning interpretations.
Meteorological reasoning and decision support Performs analysis, summarization, judgment, and decision-support reasoning based on meteorological facts, forecast information, disaster risks, and service scenarios.
Meteorological service tool calling Supports tool calling and workflow orchestration for meteorological query, data retrieval, product generation, warning services, and related tasks.
We evaluate FengHe on MetsEval-1k, a meteorological service evaluation benchmark containing 1,076 questions across 4 dimensions:
The results show that FengHe achieves a higher overall score on MetsEval-1k than other mainstream general-purpose large language models, demonstrating its domain-specific advantages in meteorological services.
python
from transformers import AutoTokenizer, AutoModelForCausalLMimport torch model_id = "PMSCCMA/FengHe" tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)model = AutoModelForCausalLM.from_pretrained( model_id, torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True,) messages = [ { "role": "user", "content": "Please generate a public-facing risk advisory for heavy rainfall." }] inputs = tokenizer.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=512, temperature=0.6, top_p=0.9,) response = tokenizer.decode( outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True,) print(response)bash
vllm serve PMSCCMA/FengHe \ --tensor-parallel-size 8 \ --served-model-name fenghebash
python3 -m sglang.launch_server \ --model-path PMSCCMA/FengHe \ --tp-size 8 \ --host 0.0.0.0 \ --port 30000 \ --served-model-name fengheCall the OpenAI-compatible API:
bash
curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "fenghe", "messages": [ { "role": "user", "content": "Please generate a meteorological service advisory for transportation under typhoon impacts." } ], "temperature": 0.6, "top_p": 0.9, "max_tokens": 512 }'FengHe is released for meteorological services, research, industry application development, and agent system construction. Users should comply with applicable laws, regulations, industry standards, and meteorological service requirements. The model must not be used to generate fake warnings, misleading meteorological information, or any content that may pose public safety risks.
This model is released under the MIT License. It can be used for research, commercial applications, and secondary development, subject to the terms of the MIT License.
from transformers import AutoTokenizer, AutoModelForCausalLMimport torch model_id = "PMSCCMA/FengHe" tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)model = AutoModelForCausalLM.from_pretrained( model_id, torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True,) messages = [ { "role": "user", "content": "Please generate a public-facing risk advisory for heavy rainfall." }] inputs = tokenizer.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=512, temperature=0.6, top_p=0.9,) response = tokenizer.decode( outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True,) print(response)vllm serve PMSCCMA/FengHe \ --tensor-parallel-size 8 \ --served-model-name fenghepython3 -m sglang.launch_server \ --model-path PMSCCMA/FengHe \ --tp-size 8 \ --host 0.0.0.0 \ --port 30000 \ --served-model-name fenghecurl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "fenghe", "messages": [ { "role": "user", "content": "Please generate a meteorological service advisory for transportation under typhoon impacts." } ], "temperature": 0.6, "top_p": 0.9, "max_tokens": 512 }'