OpenMOSS releases MOSS-VL, an 11-billion-parameter open vision-language model that changes how

models handle real-time video Instead of the standard pre-pass — ingesting every frame before the first token appears — MOSS-VL samples live frames while it generates text, pauses itself when it needs to "look," and revises its output as the scene evolves.

The architecture drops the traditional pre-pass entirely. MOSS-VL takes a live video stream, begins answering, and when it decides it needs more visual information it halts generation, grabs fresh frames, and resumes from the stopping point. OpenMOSS calls this "proactive silence," and it lets the model react to scene changes in real time rather than relying on a static representation fixed at the start of the interaction.

The complementary advance is "dynamic self-correction." When newly sampled frames contradict what the model has already said — an object moves, a person enters the frame, lighting shifts — the model detects the discrepancy and adjusts the continuation on the fly. This is not a post-hoc edit after the answer finishes; it is a mid-sentence correction, analogous to how humans re-phrase mid-utterance when they see something new.

At 11 billion parameters, MOSS-VL sits in the mid-size tier — substantially smaller than closed frontier models, yet large enough to run complex vision-language reasoning on relatively accessible hardware. The full open-weight release lets researchers and developers inspect the integrated mechanism, run ablations on the silence and correction components, and adopt the approach without depending on a closed API.

The practical upshot: "real-time video understanding" ceases to be a pre-processing step and becomes a continuous process that runs in parallel with generation. For robotics, security monitoring, or live visual assistance, the latency between perception and action collapses from seconds to fractions of a second. OpenMOSS has not published comparative benchmarks against existing models, so real-world performance still awaits independent verification.