--- title: Keye-VL-671B-A37B canonical_url: "https://www.modelscope.ai/models/Kwai-Keye/Keye-VL-671B-A37B" md_url: "https://www.modelscope.ai/models/Kwai-Keye/Keye-VL-671B-A37B.md" repository: Kwai-Keye/Keye-VL-671B-A37B last_updated: 2026-05-29 license: apache-2.0 model_type: - KeyeVL architectures: - KeyeVLMoeForConditionalGeneration parameters: 671.6B tensor_type: - F32 - F8_E4M3 - BF16 library_name: - safetensors - pytorch frameworks: - pytorch language: - en downloads: 164 stars: 0 tags: - multimodal --- # Keye-VL-671B-A37B > Keye-VL-671B-A37B - An open-source model by Kwai-Keye on ModelScope. [𝕏 X] [πŸ’¬ Discord] [🍎 Home Page] [πŸ“– Technical Report] [πŸ’» GitHub Repository] [πŸ“Š Keye-VL-8B-Preview ] [πŸ“Š Keye-VL-1.5-8B ] [πŸš€ Demo] Kwai-Keye/Keye-VL-671B-A37B is a 671.6B-parameter machine learning model on ModelScope. licensed under apache-2.0. - **Repository**: Kwai-Keye/Keye-VL-671B-A37B - **License**: apache-2.0 - **Parameters**: 671.6B - **Tags**: multimodal - **Downloads**: 164 - **Stars**: 0 - **Last updated**: 2026-05-29 Source: https://www.modelscope.ai/models/Kwai-Keye/Keye-VL-671B-A37B --- # Keye-VL-671B-A37B
Kwai Keye-VL Logo
[[𝕏 X](https://x.com/KwaiKeye)] [[πŸ’¬ Discord](https://discord.gg/4Q6AmzxpEK)] [[🍎 Home Page](https://kwai-keye.github.io/)] [[πŸ“– Technical Report](https://arxiv.org/abs/2509.01563)] [[πŸ’» GitHub Repository](https://github.com/Kwai-Keye/Keye)] [[πŸ“Š Keye-VL-8B-Preview](https://huggingface.co/Kwai-Keye/Keye-VL-8B-Preview) ] [[πŸ“Š Keye-VL-1.5-8B](https://huggingface.co/Kwai-Keye/Keye-VL-1_5-8B/) ] [[πŸš€ Demo](https://huggingface.co/spaces/Kwai-Keye/Keye-VL-8B-Preview)]
Meet Keye-VL-671B-A37B β€” the most powerful multi-modal language model in the Keye series to date. As one of the largest and most capable MLLMs currently in existence, Keye-VL-671B-A37B demonstrates top-tier and in some cases even leading performance in text understanding and generation, complex visual perception and reasoning, comprehensive video understanding, and Olympic-level mathematical reasoning. #### Key Enhancements: ##### Pre-Training * **Efficient Perception Building with Limited Compute**: We employ VisionEncoder from Keye-VL-1.5 and rigorously processed high-quality data to cost-effectively build the model’s core perceptual capabilities, ensuring strong visual understanding without excessive computational overhead. * **Multi-Modal Data Curation**: We implement a automated data pipeline that performs strict filtering, re-sampling, and large-scale synthesis of structured VQA data, including OCR, charts, and tables. This end-to-end process significantly enhances the model’s perception quality and generalization. * **Reasoning Sustainment via Synthetic CoT Data**: During the continual pretrain phase, we incorporate a diverse set of synthetically generated chain-of-thought (CoT) data. This ensures the model maintains its complex reasoning skills while progressing in perceptual pre-training. ##### Post-Training * **Scaling law of Reasoning Data for SFT**: We experimentally validate that the mixed data (50B Instruct & Long-CoT data) improves model performance and training stability compared to the single model (30B Instruct data). * **CoT Quality & Style Refinement**: We develop a data filtering process to remove redundant reflective chains, improving the model's reasoning and perception capabilities, with the in-house process outperforming GPT-4o. * **High-Precision RL Verifier**: We train a dedicated verifier (Keye Verifier) to validate the model's reasoning consistency and answer correctness, achieving significantly higher accuracy than other reward models and general LLMs, thereby enhancing our RL performance gains. ## Model Performance ![Performance Comparison](figures/radar.png) ![Performance on Public Benchmarks](figures/performance.png) ## Quickstart ### Environment Setup ```shell docker run -it --gpus all lmsysorg/sglang:v0.5.2 # make sure each node use the following commands to install the custom SGLang branch git clone -b keye-dpsk-infer-fp8-release https://github.com/Kwai-Keye/sglang.git sglang pip install -e sglang/python[all] ``` ### Two-Node H800x8 Deployment #### Prerequisites: - Model: Kwai-Keye/Keye-VL-671B-A37B - Node 1 IP: 192.168.1.100 (MASTER_NODE_IP) - Node 2 IP: 192.168.1.101 (WORKER_NODE_IP) #### Node 1 (Master - rank 0): ```shell MODEL_PATH=/path/to/Keye-VL-671B-A37B DIST_INIT_ADDR="MASTER_NODE_IP:29500" # e.g. 192.168.1.100:29500 PORT=30000 # listening port on each node python3 -m sglang.launch_server \ --model-path $MODEL_PATH \ --host 0.0.0.0 \ --port $PORT \ --tp-size 16 \ --nnodes 2 \ --node-rank 0 \ --dist-init-addr $DIST_INIT_ADDR \ --trust-remote-code \ --mm-attention-backend fa3 \ --attention-backend fa3 \ --disable-radix-cache \ --mem-fraction-static 0.8 \ --cuda-graph-max-bs 64 \ --model-loader-extra-config '{"enable_multithread_load": true, "num_threads": 32}' ``` #### Node 2 (Worker - rank 1): ```shell MODEL_PATH=/path/to/Keye-VL-671B-A37B DIST_INIT_ADDR="MASTER_NODE_IP:29500" # e.g. 192.168.1.100:29500 PORT=30000 # listening port on each node python3 -m sglang.launch_server \ --model-path $MODEL_PATH \ --host 0.0.0.0 \ --port $PORT \ --tp-size 16 \ --nnodes 2 \ --node-rank 1 \ --dist-init-addr $DIST_INIT_ADDR \ --trust-remote-code \ --mm-attention-backend fa3 \ --attention-backend fa3 \ --disable-radix-cache \ --mem-fraction-static 0.8 \ --cuda-graph-max-bs 64 \ --model-loader-extra-config '{"enable_multithread_load": true, "num_threads": 32}' ``` For more deployment details, please refer to the [Keye-VL-671B-A37B Deployment Tutorial](https://github.com/Kwai-Keye/sglang/blob/keye-dpsk-infer-fp8-release/scripts/deploy_keye_deepseek/DEPLOY_TUTORIAL.md). ### Client Usage ```python import json import requests BASE_URL = "http://MASTER_NODE_IP:30000" def generate(messages): payload = { "model": "", "messages": messages, "n": 1, "temperature": 0.0, "max_tokens": 256, "top_k": 1, "ignore_eos": False, "skip_special_tokens": True, } resp = requests.post( f"{BASE_URL}/v1/chat/completions", headers={"Content-Type": "application/json"}, data=json.dumps(payload), timeout=1800, ) resp.raise_for_status() return resp.json() # Example: image + text messages = [ { "role": "user", "content": [ { "type": "image_url", "image_url": {"url": "https://raw.githubusercontent.com/sgl-project/sglang/main/assets/logo.png"}, }, {"type": "text", "text": "Describe this image in detail./think"}, ], } ] result = generate(messages) print(result["choices"][0]["message"]["content"]) ```