Agnes Video 2.0 MoE ships open weights under Apache 2.0

Agnes-Video-2.0-MoE, a 28.7-billion-parameter video model that also generates audio, ships its weights under the Apache License 2.0.

Quantslant reported the upload on October 11, 2026, and identifies the maker as Singapore-based Sapiens AI. The Hugging Face repository is live, holding the model card, an Apache 2.0 LICENSE file, and seven weight shards that total about 57.5 GB.

The checkpoint is an audio-video generation model built on DaVinci-MagiHuman, the open project from the GAIR-NLP team whose repository is also Apache 2.0. It accepts text, video and audio as input and outputs video with audio. The model card describes a 40-layer single-stream DiT with three-expert SparseMoE blocks and Top-1 routing on layers 4 to 35, at 28.721 billion stored parameters and 15.299 billion active.

The Apache 2.0 label covers less than "open weights" suggests. The model card says the Agnes code and weights are provided under the Apache License 2.0, subject to applicable upstream license requirements, and that the license does not extend to external components. The T5Gemma 9B text encoder is governed by the Gemma Terms of Use, not Apache 2.0, and has to be obtained separately. The video VAE comes from Wan2.2-TI2V-5B, and the audio encoder is OmniVAE. The repository's NOTICE file states that the release redistributes none of those components, nor any DaVinci-MagiHuman base checkpoint.

The open release arrives as Agnes AI moves its video line to a closed, paid version. The company retired Agnes-Video-V2.0 from its API on September 25, 2026, according to its retirement notice, and its pricing page now lists only Agnes Video 2.5 and 2.5 Flash. Agnes Video 2.5 is not open weights. The page bills 720p output at $0.025 per second and 1080p at $0.040 per second, or $0.20 for a 5-second 1080p clip.

Self-hosting the full model is a server job, not a desktop one. The shards total about 57.5 GB, more than double the 24 GB on an RTX 4090, and the model card's benchmarks run at 448x256 for 4 to 5 seconds on an NVIDIA H100 in FP32. The card also promises an "EENP 60%" pruned checkpoint that would cut DiT denoising latency from 26.83 seconds to 20.92 seconds, while its VBench 2.0 score drops from 0.519 to 0.479; those weights are not uploaded yet. The free hosted tier is strained in the meantime: a GitHub issue opened October 10 reports sustained 503 "video_queue_full" errors on agnes-video-2.5-flash, and an Agnes team account replied on October 12 that server queue capacity is limited.

The unpruned weights and the inference script are on Hugging Face now; the T5Gemma encoder, OmniVAE, and the Wan2.2 VAE have to be pulled separately from their own repositories.