MOSS-VL 技術報告

摘要

我們提出 MOSS-VL，一個開放原始碼的視覺-語言模型系列，將即時互動——在說話的同時進行感知——視為一項首要能力。此模型在整個技術棧上進行協同設計：語言解碼器僅透過門控交叉注意力機制關注視覺訊息，使模型在生成時能自然地感知輸入畫面；一個合成的互動語料庫監督何時說話、何時保持沉默、以及何時修正；並採用分階段課程，將所有即時相關的訓練集中於一個輕量的最終階段，建立於強大的離線基礎之上。在離線設定中，MOSS-VL-Instruct 在相近規模下具有競爭力，並在時間推理影片集上取得領先。在四個串流基準測試中，MOSS-VL-Realtime 於開源串流模型中，在三個基準上取得最佳平均成績（第四個基準名列第二），並橫掃了直接測試主動行為的三個子集——在 OmniMMI Proactive Alerting 上取得 66.0，而最佳基線僅為 37.5。憑藉 113 億參數，且視覺 token 位於解碼序列之外，MOSS-VL 在視覺上下文增長時，將其相對於同骨幹架構之 Qwen3-VL-8B 的首 token 延遲優勢從 2.8 倍擴大至 5.1 倍。我們已於 https://github.com/OpenMOSS/MOSS-VL 釋出全部五個檢查點、訓練課程，以及即時推論程式碼。

English

We present MOSS-VL, an open vision-language model family that treats real-time interaction -- perceiving while it speaks -- as a first-class capability. It is co-designed across the stack: the language decoder attends to vision only through gated cross-attention, so the model can naturally see incoming frames while generating; a synthesized interaction corpus supervises when to speak, when to stay silent, and when to revise; and a staged curriculum concentrates all real-time-specific training in one light final stage over a strong offline foundation. Offline, MOSS-VL-Instruct is competitive at comparable scale and leads temporal-reasoning video sets. Across four streaming benchmarks, MOSS-VL-Realtime posts the best average on three (second on the fourth) among open-source streaming models, sweeping the three subsets that squarely test proactive behavior -- 66.0 vs. 37.5 for the best baseline on OmniMMI Proactive Alerting. With 11.3B parameters but visual tokens outside the decoded sequence, MOSS-VL widens its time-to-first-token advantage over same-backbone Qwen3-VL-8B from 2.8x to 5.1x as visual context grows. We release all five checkpoints, the training curriculum, and the real-time inference code at https://github.com/OpenMOSS/MOSS-VL.