MOSS-Video-Preview¶

Development line: project:moss-video-preview · thread moss-video-preview
Last event: 2026-06-10 · 1 dated since 2026-06-10 · Researched: 2026-09-05 · confidence: medium

What it is¶

MOSS-Video-Preview is an Apache-2.0, custom-code video-language model family built on gated image-video cross-attention: Base for pretraining, SFT for offline instruction following, and Real-Time-SFT for live frame-by-frame inference. The released Base checkpoint is 11B BF16; the project reports a single-H200 measurement of 1.9537 s average TTFT on a 256-frame test. Verdict: use Real-Time-SFT only when streaming behavior is required; this is an exploratory research release, not a hosted inference service.

Development line¶

• 2026-06-10 — MOSS-Video-Preview public project resources referenced. On 2026-06-10, the MOSS-Video-Preview development line was associated with a public GitHub repository and a Hugging Face collection. The dated links establish a public project reference point, but do not by themselves establish a specific release, model capability, or version change.

What changed¶

2026-06-10 — The linked project package was publicly discoverable; first-party material identifies the family as a roughly 10.7B/11B gated-cross-attention video model with separate Base, offline SFT, and Real-Time-SFT checkpoints. 2026-06-01 — OpenMOSS published the technical report and released Realtime-QA-100K, a 100,000-sample training subset for real-time video QA. 2026-04-08 — OpenMOSS released MOSS-VL-Base-0408 and MOSS-VL-Instruct-0408 after starting the MOSS-VL line. 2026-03-04 — Source code and architecture details for MOSS-Video-Preview were released. 2025-10-18 — The project recorded a post-mortem and began MOSS-VL. 2025-10-08 — An internal demo was shown. 2025-09 — The Real-Time-SFT checkpoint was ready. 2025-08 — The offline SFT checkpoint was ready.

How to use this¶

As of 2026-06-10, practitioners should use the linked GitHub repository and Hugging Face collection as the initial public reference points for MOSS-Video-Preview, while verifying any claimed release details from primary project materials.

1. Clone the repository, create the documented Python 3.12.4 environment, and install the package; the tested stack is PyTorch 2.4.0 with CUDA 12.1 and DeepSpeed 0.16.1. — https://github.com/OpenMOSS/MOSS-Video-Preview
2. Use moss-video-preview-sft with inference.offline_infer for a prerecorded video and an instruction prompt; use the Base checkpoint for completion or as an SFT starting point. — https://github.com/OpenMOSS/MOSS-Video-Preview
3. Use moss-video-preview-realtime-sft with inference.realtime_streaming_infer for incoming frames; Base and plain SFT are not compatible with the streaming path. — https://github.com/OpenMOSS/MOSS-Video-Preview
4. Load Realtime-QA-100K with the Hugging Face datasets library when reproducing or adapting real-time SFT data; it provides annotations, YouTube IDs, and timestamps rather than video files. — https://huggingface.co/datasets/OpenMOSS-Team/Realtime-QA-100K

Best practices¶

• Select the checkpoint by task: Base needs downstream SFT for instruction following, offline SFT is for whole clips, and Real-Time-SFT is the required checkpoint for streaming. — https://huggingface.co/OpenMOSS-Team/moss-video-preview-base
• Use trust_remote_code=True only after reviewing the repository code, because this model family requires custom Transformers mappings; use FlashAttention 2 on CUDA when low latency matters. — https://huggingface.co/OpenMOSS-Team/moss-video-preview-realtime-sft
• Treat the published speed comparison as one H200 setup, not as a general benchmark; its test used 256 extracted frames from one 97.56-second, 1080p video. — https://github.com/OpenMOSS/MOSS-Video-Preview
• For Realtime-QA-100K, plan for video availability and rights yourself: the release contains no videos, and the original YouTube material can disappear or be region-restricted. — https://huggingface.co/datasets/OpenMOSS-Team/Realtime-QA-100K

Superseded by this¶

• Nothing marked obsolete yet.

Still unknown¶

• The supplied Hugging Face collection endpoint did not return readable content during this check, so its creation date and the exact collection state on 2026-06-10 are unverified.
• No first-party source examined declares MOSS-Video-Preview formally deprecated or states that MOSS-VL supersedes it; MOSS-VL is therefore recorded as a related successor development line, not a formal replacement.
• The project’s current README still describes further scaling, distributed training, and broader code/data releases as future work; production readiness, maintained release cadence, and third-party serving support are unverified.
• event_findings:[2026-06-10] The project documentation identifies the linked release as an approximately 10.7B (marketed as 11B) model family with 40 decoder layers, including 8 gated cross-attention layers; it separates Base, offline SFT, and Real-Time-SFT rather than shipping one interchangeable checkpoint. Source date: 2026-06-01 project news entry; source URL: https://github.com/OpenMOSS/MOSS-Video-Preview.

Sources¶

Agent brief¶

• Subject: project:moss-video-preview, thread moss-video-preview, 1 dated events 2026-06-10 → 2026-06-10.
• Practical note: As of 2026-06-10, practitioners should use the linked GitHub repository and Hugging Face collection as the initial public reference points for MOSS-Video-Preview, while verifying any claimed release details from primary project materials.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.