Hugging Face Daily Papers · August 18, 2026 · 4 min read

MOSS-VL Technical Report

Mirrored from Hugging Face Daily Papers for archival readability. Support the source by reading on the original site.

Like Read original ↗

MOSS-VL treats real-time interaction — perceiving while it speaks — as a first-class capability: vision enters the decoder only through gated cross-attention, so the model keeps watching incoming frames while it generates. Among open-source streaming models it posts the best average on 3 of 4 streaming benchmarks and sweeps all three proactive subsets (66.0 vs. 37.5 on OmniMMI Proactive Alerting), while widening its TTFT advantage over same-backbone Qwen3-VL-8B from 2.8× to 5.1× as visual context grows. All five checkpoints, the training curriculum, and the real-time inference code are released.</p>\n","updatedAt":"2026-08-18T09:06:47.340Z","author":{"_id":"62a148a815eeab266b2cd3ed","avatarUrl":"/avatars/bd6dfa112dbd98d75046789f52073e3a.svg","fullname":"Pengyu Wang","name":"Jihuai","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":10,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.9114390015602112},"editors":["Jihuai"],"editorAvatarUrls":["/avatars/bd6dfa112dbd98d75046789f52073e3a.svg"],"reactions":[{"reaction":"🔥","users":["ShacklesLay"],"count":1}],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2608.15045","authors":[{"_id":"6a84099eb153becad1677028","name":"Pengyu Wang","hidden":false},{"_id":"6a84099eb153becad1677029","name":"Chenkun Tan","hidden":false},{"_id":"6a84099eb153becad167702a","name":"Shaojun Zhou","hidden":false},{"_id":"6a84099eb153becad167702b","name":"Qirui Zhou","hidden":false},{"_id":"6a84099eb153becad167702c","name":"Yanxin Chen","hidden":false},{"_id":"6a84099eb153becad167702d","name":"Xingyang He","hidden":false},{"_id":"6a84099eb153becad167702e","name":"Huazheng Zeng","hidden":false},{"_id":"6a84099eb153becad167702f","name":"Jijun Cheng","hidden":false},{"_id":"6a84099eb153becad1677030","name":"Chenghao Wang","hidden":false},{"_id":"6a84099eb153becad1677031","name":"Xiaomeng Qian","hidden":false},{"_id":"6a84099eb153becad1677032","name":"Pengfei Wang","hidden":false},{"_id":"6a84099eb153becad1677033","name":"Zhan Huang","hidden":false},{"_id":"6a84099eb153becad1677034","name":"Shanqing Gao","hidden":false},{"_id":"6a84099eb153becad1677035","name":"Wei Huang","hidden":false},{"_id":"6a84099eb153becad1677036","name":"Longjun Cao","hidden":false},{"_id":"6a84099eb153becad1677037","name":"Wu Ran","hidden":false},{"_id":"6a84099eb153becad1677038","name":"Jie Liu","hidden":false},{"_id":"6a84099eb153becad1677039","name":"Changtai Zhu","hidden":false},{"_id":"6a84099eb153becad167703a","name":"Hongkai Wang","hidden":false},{"_id":"6a84099eb153becad167703b","name":"Yixian Tian","hidden":false},{"_id":"6a84099eb153becad167703c","name":"Chenghao Liu","hidden":false},{"_id":"6a84099eb153becad167703d","name":"Zhen Ye","hidden":false},{"_id":"6a84099eb153becad167703e","name":"Xinghao Wang","hidden":false},{"_id":"6a84099eb153becad167703f","name":"Botian Jiang","hidden":false},{"_id":"6a84099eb153becad1677040","name":"Guoguo Feng","hidden":false},{"_id":"6a84099eb153becad1677041","name":"Zhaoye Fei","hidden":false},{"_id":"6a84099eb153becad1677042","name":"Ruixiao Li","hidden":false},{"_id":"6a84099eb153becad1677043","name":"Mingshu Chen","hidden":false},{"_id":"6a84099eb153becad1677044","name":"Yang Gao","hidden":false},{"_id":"6a84099eb153becad1677045","name":"Qinyuan Cheng","hidden":false},{"_id":"6a84099eb153becad1677046","name":"Shimin Li","hidden":false},{"_id":"6a84099eb153becad1677047","name":"Xipeng Qiu","hidden":false}],"publishedAt":"2026-08-15T00:00:00.000Z","submittedOnDailyAt":"2026-08-18T00:00:00.000Z","title":"MOSS-VL Technical Report","submittedOnDailyBy":{"_id":"62a148a815eeab266b2cd3ed","avatarUrl":"/avatars/bd6dfa112dbd98d75046789f52073e3a.svg","isPro":false,"fullname":"Pengyu Wang","user":"Jihuai","type":"user","name":"Jihuai"},"summary":"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.","upvotes":32,"discussionId":"6a84099eb153becad1677048","projectPage":"https://openmoss.ai/MOSS-VL/","githubRepo":"https://github.com/OpenMOSS/MOSS-VL","githubRepoAddedBy":"user","ai_summary":"MOSS-VL is an open vision-language model family enabling real-time interaction by attending to vision via gated cross-attention during generation, using a synthesized interaction corpus and staged curriculum to achieve strong streaming performance with reduced time-to-first-token latency.","ai_keywords":["vision-language model","gated cross-attention","real-time interaction","streaming benchmarks","time-to-first-token","temporal reasoning","interaction corpus","staged curriculum"],"ai_summary_model":"thinkingmachines/Inkling-Small","githubStars":430,"organization":{"_id":"613b0dee83ec35d460684607","name":"OpenMOSS-Team","fullname":"OpenMOSS","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/61457b8deff2c9fdb4de4988/N5b9663zQ4uq5_OTNlnmw.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"62a148a815eeab266b2cd3ed","avatarUrl":"/avatars/bd6dfa112dbd98d75046789f52073e3a.svg","isPro":false,"fullname":"Pengyu Wang","user":"Jihuai","type":"user"},{"_id":"68b132f80ff21ef5993997d1","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/OGo_F891RFaTBykgJPrkg.png","isPro":false,"fullname":"CHEN Yanxin","user":"CCCCyx","type":"user"},{"_id":"6594ffe3a78a277803814a99","avatarUrl":"/avatars/565f30793ff2afe6656af275d57773a5.svg","isPro":false,"fullname":"Shaojun Zhou","user":"sjzhou","type":"user"},{"_id":"644a41fcd9a3ae8341055179","avatarUrl":"/avatars/21b35abdc60a34589443b5879901eb46.svg","isPro":false,"fullname":"Changtai Zhu","user":"BeastyZ","type":"user"},{"_id":"66cc484f315af068a9b583d2","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/66cc484f315af068a9b583d2/8rHQckULHPIgekH2DXSD2.jpeg","isPro":false,"fullname":"Chen","user":"Qiancccc","type":"user"},{"_id":"6612836ee60b8224786bc54f","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6612836ee60b8224786bc54f/Bvs1KQqBmZzDvkSvjEwrX.png","isPro":false,"fullname":"Chenghao Wang","user":"henrywch2huggingface","type":"user"},{"_id":"677e4069ad21b67f2c8783d2","avatarUrl":"/avatars/9b31108eec388a870fba72fb9b83c7f9.svg","isPro":false,"fullname":"zhou","user":"ephurii","type":"user"},{"_id":"6899510f81115b07e64f4e63","avatarUrl":"/avatars/b38adec896dc651b67afca3ba0388506.svg","isPro":false,"fullname":"GaoXingJian","user":"Bdyskov","type":"user"},{"_id":"64c3c631e77ea9f28111172a","avatarUrl":"/avatars/495dbb73b69c399bae780da3118e332f.svg","isPro":false,"fullname":"Siyin Wang (SII)","user":"sinwang","type":"user"},{"_id":"65acd3c60844d9e0d6a0ee27","avatarUrl":"/avatars/73c3f46412f11db371e277442aa83d9d.svg","isPro":false,"fullname":"qi","user":"ning777","type":"user"},{"_id":"680f7d6b8b2e2c7db910962c","avatarUrl":"/avatars/99d78c3c7d04a79121409c10d84df83a.svg","isPro":false,"fullname":"huazzeng","user":"huazzeng","type":"user"},{"_id":"6a62dd6277610717cca52d6d","avatarUrl":"/avatars/0985530cdf0abb88e2738ed7f3872911.svg","isPro":false,"fullname":"zz","user":"zz913","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"613b0dee83ec35d460684607","name":"OpenMOSS-Team","fullname":"OpenMOSS","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/61457b8deff2c9fdb4de4988/N5b9663zQ4uq5_OTNlnmw.png"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2608/2608.15045.md","query":{}}">

Papers

arxiv:2608.15045

MOSS-VL Technical Report

Published on Aug 15

· Submitted by

Pengyu Wang on Aug 18

OpenMOSS

Upvote

Authors:

Abstract

MOSS-VL is an open vision-language model family enabling real-time interaction by attending to vision via gated cross-attention during generation, using a synthesized interaction corpus and staged curriculum to achieve strong streaming performance with reduced time-to-first-token latency.

Generated by thinkingmachines/Inkling-Small

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.