Wan 3.0 Ecosystem Signals: What WanSong, Wan-Dancer, and Wan-Streamer Reveal About Alibaba's Next Video Model
Video model season is about to start.
Hailuo 3 out tomorrow (July 29). Flux 3 public release August 4. WAN 3 release August 6.
Still waiting on Seedance 2.5 date.
— @koltregaskes, h/t MarsForTech
If the rumor holds, Wan 3.0 drops on August 6, 2026 — less than two weeks from today. It would cap a July that saw three architectural previews (WanSong, Wan-Dancer, Wan-Streamer), a new FLUX 3 announcement from Black Forest Labs, and the Kimi K3 open-weight release. The open-source video generation ecosystem has never moved this fast.
But rumors aside, the technical evidence for Wan 3.0's direction is already visible in the published papers and released weights. Between July 10 and July 16, 2026, Alibaba's Wan team published technical reports on WanSong (text-to-music, ~25B parameters) and Wan-Streamer v0.3 (real-time streaming interaction), and released Wan-Dancer-14B (music-to-dance) under Apache 2.0. None carries a "Wan 2.x" label. All three operate across modalities that Wan 2.2 never touched — audio, motion, real-time interaction.
This article examines what each release reveals about the shared backbone these models build on, what a potential Wan 3.0 architecture might look like, and what practitioners can do today. We focus on what is independently verifiable: the papers on arXiv, the model weights on HuggingFace, and the model cards — not just the leaks.
1. The Wan Ecosystem Before Wan 3.0
1.1 Current Public Models
| Model | Release | Type | Parameters | Notes |
|---|---|---|---|---|
| Wan 2.1 | Feb 2025 | T2V / I2V | 1.3B, 14B | First open-source Wan release |
| Wan 2.2 | Jul–Aug 2025 | T2V, I2V, S2V, Animate | 5B–17B | Video-only DiT with flow matching |
| Wan-Dancer-14B | Jul 2026 | Music-to-Dance | 14B | Built on Wan-I2V (Apache 2.0) |
| WanSong v1.0 | Jul 2026 | Text-to-Music | ~25B | Paper only, no weights yet |
| Wan-Streamer v0.3 | Jul 2026 | Real-time AV streaming | Undisclosed | Paper only |
Wan 2.2 was the last release to carry the "2.x" branding. Everything since — Dancer, Song, Streamer — uses distinct names and versioning. They share architectural DNA but are not labeled as Wan 2.x extensions. The question is whether they represent components being developed for Wan 3.0, or independent research directions.
1.2 What Changed After Wan 2.2
Wan 2.2 is a pure video generation pipeline — text-to-video, image-to-video, speech-to-video, video-to-video animation. Its architecture is a Diffusion Transformer (DiT) with flow matching, a 3D VAE for video compression, and a T5 text encoder for conditioning.
The three post-2.2 releases operate across modalities (audio, video, text, motion), use multimodal backbones for multiple downstream tasks, and in the case of Wan-Streamer, handle real-time streaming interaction — a fundamentally different serving paradigm from Wan 2.2's batch-generation pipeline.
The architectural answers are scattered across three papers published within 72 hours of each other.
2. Three Releases That Signal Wan 3.0's Architecture
2.1 WanSong v1.0: A ~25B Pure Diffusion Music Model
WanSong (July 16, 2026) generates up to 5 minutes of multilingual songs with separate vocal and background music stems in a single forward pass. The paper describes it as having "nearly 25B" parameters.
Architecture:
- Uses a hybrid-MMDiT backbone — the same design pattern introduced by FLUX. Instead of separate parameters for each modality in every layer, it uses shared AdaLN blocks, "following our previous settings in Wan2.1" (Section 3.2).
- Text tokens (from a decoder-only LLM captioner) and dual-stem audio continuous tokens (from a 1D VAE) are packed into a single unified sequence.
- Vocal and BGM tokens are produced independently at the model output but learned jointly within each block. This solves the CFG tradeoff where higher guidance improves vocals but suppresses BGM.
Audio VAE:
- 1D continuous VAE at 44.1 kHz stereo with 1024× compression (vs. Stable Audio 2's 2048×)
- SI-SDR: 7.25 dB on music (vs. SA2's 4.39 dB), 12.92 dB on speech (vs. SA2's 8.65 dB)
- Training data: ~260M audio clips (50% music, 40% speech, 10% sound)
Training data: "More than 6 million hours of multilingual song data" with a 5-stage pipeline. Captions include musical style, vocals, instruments, mode and key, language, reverb, emotions, and lyrics.
RLHF pipeline: Two-stage alignment — DPO for global structure, then ReFL for fine-grained detail in the low-noise regime.
Benchmark highlights:
| Metric | WanSong | Suno V5.5 | Suno V5 |
|---|---|---|---|
| Melody | 7.02 | 6.76 | 6.98 |
| Arrangement | 7.25 | 7.05 | 7.15 |
| Vocal | 7.38 | 6.83 | 7.25 |
| PER (lyric accuracy) ↓ | 7.43% | 9.86% | 22.80% |
Wan 3.0 signal: WanSong proves the Wan team has built a multimodal DiT backbone — the hybrid-MMDiT — that handles audio natively, not as an afterthought. The shared AdaLN design traces back to Wan 2.1, meaning the multimodal architecture was planned for extensibility from the start. If Wan 3.0 generates synchronized audio with video, this is the backbone it will build on.
2.2 Wan-Dancer-14B: Hierarchical Video Generation for Wan 3.0
Wan-Dancer (July 10, 2026, revised July 17) generates minute-scale dance videos from music. It builds on Wan-I2V as its backbone (a model in the Wan 2.x family). The model is available on HuggingFace under Apache 2.0.
The official announcement from @Alibaba_Wan on July 14 confirmed the release, and the community response was immediate — @ModelScope highlighted the breakthrough beyond the 20-second video barrier with 720p/30fps output across five dance genres. @wildmindai noted the 85GB model size, while Chinese AI commentator @berryxia emphasized the significance of local deployment: "不再是又一个通用视频模型，而是把'音乐→节奏→精准舞蹈动作'这个垂直能力做到了可本地部署的水平" (no longer just another general video model — it brings the vertical capability of music-to-precise-dance-motion to local deployability).
Two-stage hierarchical generation:
- Global stage: Generates sparse keyframes (38 frames) spanning the full music track at low resolution, using full-track musical context for long-range coherence. The keyframe mask forces the model to infer structural evolution from the initial frame alone.
- Local stage: Refines each 5-second segment (149 frames) around the keyframes at 720p/30fps. Local segments run in parallel.
- Dynamic frame rate: Absolute time identifiers are integrated into RoPE, supporting 3–15 fps.
Inference parameters (from the model card):
- Global stage: 48 steps, CFG scale = 5.0
- Local stage: 24 steps, CFG scale = 5.0
Results:
| Metric | MusicInfuser | X-Dancer | Wan-Dancer |
|---|---|---|---|
| Dance Quality | 6.23 | 6.06 | 8.46 |
| Video Quality | 5.22 | 6.23 | 7.46 |
| Prompt Alignment | 6.61 | — | 9.03 |
Supports five dance genres: Chinese Classical, K-Pop, Latin, Tap, Street. Training used 128 A100 GPUs.
# Inference with Wan-Dancer via Diffusers
import torch
from diffusers import WanDancerPipeline
pipe = WanDancerPipeline.from_pretrained(
"Wan-AI/Wan-Dancer-14B",
torch_dtype=torch.bfloat16
).to("cuda")
video = pipe(
image="dancer_ref.png",
audio="music_input.wav",
prompt="K-pop dance with precise beat synchronization",
num_inference_steps_global=48,
num_inference_steps_local=24,
cfg_scale=5.0,
).video[0]
Hardware note: At 14B parameters in bfloat16 (~28 GB VRAM minimum), this sits at the edge of consumer GPU range. An RTX 5090 (32 GB) may suffice with optimization; an A6000 48GB or A100 80GB is safer for production.
Wan 3.0 signal: Two-stage hierarchical generation is a capability Wan 2.2 does not have. If Wan 3.0 adopts this paradigm, it would support much longer video — potentially minutes rather than seconds — with global coherence that single-pass models cannot achieve. Wan-Dancer's music conditioning pipeline also suggests Wan 3.0's audio conditioning is richer than anything in Wan 2.x.
2.3 Wan-Streamer v0.3: The Pretraining Paradigm for Wan 3.0
Wan-Streamer v0.3 (July 16, 2026) reframes video generation as a world + event stream decomposition — a general-purpose pretraining objective, not a specific model architecture. The paper is explicit: "this is not a new architecture: it is a reading of the same interleaved causal sequence" (Section 2.1).
The decomposition: A video is a persistent world (environment, scene, subjects, ambient acoustics, voice characteristics) plus an event stream (everything that changes — behavior, speech, sounds). The model learns to predict how a world streams forward given an input, unit by unit.
Key specifications:
- 640×368 at 25 FPS, 160 ms streaming unit, ~200 ms model-side latency
- Total interaction latency: ~550 ms (including 350 ms network budget)
- Full-duplex communication — both parties can speak and move simultaneously
- Block-causal attention over language, audio, and video inputs/outputs
- Open-vocabulary behavior actions via parenthesized directives:
"(reaches into the grass and picks up a green leaf) Look what I found." - Serving topology inherited from v0.2 (Thinker + Ulysses-style context-parallel performer)
The code is CC BY 4.0. No model weights released.
The v0.3 release drew significant attention. @minchoi called it "not voice mode anymore 🤯" with over 3,400 likes, and Japanese AI analyst @itarutomy provided a detailed technical breakdown, highlighting the paradigm shift from v0.1/v0.2's agent-specific training to v0.3's general-purpose world-event pretraining followed by downstream specialization. @ai_hakase_ emphasized the full-duplex capability — natural conversational interruptions and simultaneous speech — as the key breakthrough.
Wan 3.0 signal: This is the most architecturally significant of the three releases. The world-event decomposition is a general-purpose pretraining objective that can be applied to any causal transformer — including Wan 3.0's backbone. If Wan 3.0 adopts this framing, it's no longer just a video generator. It becomes a general-purpose world model that can be specialized to any task involving real-time prediction of how scenes evolve. That's a category jump — from "text-to-video model" to "multimodal world simulator."
3. The Wan 3.0 Architecture: What the Evidence Points To
3.1 Shared Backbone Evidence
All three models point toward a common architectural family. Here's what we can confirm from the published materials:
| Component | Wan 2.2 | Wan 3.0 Direction (from releases) |
|---|---|---|
| Backbone | DiT (video only) | Hybrid-MMDiT (text + audio + video) |
| Video tokenizer | 3D VAE | 3D VAE (same or improved) |
| Audio tokenizer | None | 1D VAE (1024×, 44.1 kHz) |
| Text encoder | T5 | LLM-based captioner |
| Generation paradigm | Single-pass flow matching | Hierarchical two-stage or world-event streaming |
| Max video length | ~5 seconds | Minutes (via hierarchical planning) |
| Real-time mode | No | 25 FPS, ~200ms latency |
| Output modalities | Video | Video + audio + behavior |
Evidence 1: WanSong explicitly follows Wan 2.1's AdaLN design. Section 3.2 states "following our previous settings in Wan2.1, we employ the fully-shared AdaLN in each block." The multimodal backbone was designed with extensibility in mind from Wan 2.1 onward.
Evidence 2: Wan-Streamer's decomposition works on any causal transformer. The paper states the world-event framing is "not a new architecture" — it's a training objective that can be applied to Wan 3.0's backbone regardless of its specific architecture.
Evidence 3: Wan-Dancer builds on existing Wan video weights. It inherits video capabilities from Wan-I2V rather than training from scratch, suggesting a modular design where Wan 3.0 capabilities can be stacked incrementally.
3.2 What Wan 3.0 Capabilities Would Look Like
If Wan 3.0 ships with the combined capabilities of these releases, the feature set would include:
| Capability | Source | Status |
|---|---|---|
| Text-to-Video | Wan 2.2 heritage | Production-ready |
| Image-to-Video | Wan 2.2 heritage | Production-ready |
| Speech-to-Video | Wan 2.2 S2V | Production-ready |
| Video Animation | Wan 2.2 Animate | Production-ready |
| Text-to-Music (5 min, dual-stem) | WanSong | Research stage |
| Music-to-Dance (minute-scale) | Wan-Dancer | Apache 2.0 release |
| Real-time AV streaming interaction | Wan-Streamer | Research stage |
| Open-vocabulary behavior generation | Wan-Streamer v0.3 | Research stage |
| Hierarchical long-video generation | Wan-Dancer | Apache 2.0 release |
3.3 What We Cannot Confirm About Wan 3.0
That Wan 3.0 is a planned product. Alibaba has not announced a Wan 3.0 release date, model card, or technical report. The architectural convergence is real, but these could be independent research directions rather than components of a unified release.
That these models share a single set of weights. WanSong (~25B) and Wan-Dancer (14B) operate at different scales with different architectures. Wan-Streamer's size is undisclosed. A unified Wan 3.0 would need to reconcile these differences.
The parameter scale of a unified model. WanSong is 25B for audio-only. A unified video+audio model would need additional parameters for video encoding/decoding and cross-modal attention. Budget for 30B+ as a working estimate, not a specification.
That Wan-Streamer's latency generalizes. Wan-Streamer's ~200ms is at 640×368 with a specialized streaming topology. A unified Wan 3.0 doing text-to-video at higher resolutions with synchronized audio will have different latency characteristics.
The license. Wan 2.x and Wan-Dancer are Apache 2.0. WanSong and Wan-Streamer have no published license. Do not ship a product that depends on Wan 3.0 being Apache 2.0.
4. Release Timeline Signals
An unconfirmed leak places Wan 3.0 on August 6, 2026 — a date that would align Hailuo 3 (July 29), FLUX 3 public release (August 4), and Wan 3.0 in a single "video model season." The source, @koltregaskes citing MarsForTech, has not been verified by Alibaba.
Regardless of the exact date, the cluster of official releases is already notable:
- July 10, 2026: Wan-Dancer paper on arXiv
- July 13, 2026: Wan-Dancer-14B released on HuggingFace (Apache 2.0)
- July 16, 2026: WanSong and Wan-Streamer v0.3 papers published
- July 14, 2026: Official Wan-Dancer announcement from @Alibaba_Wan
Three major publications within one week suggests a coordinated push. WanSong is at "v1.0," Wan-Streamer at "v0.3" — both versioned as if they are components converging toward a major release. A Wan 3.0 announcement in Q3–Q4 2026 is a reasonable working hypothesis, but it remains a hypothesis until Alibaba confirms it.
The broader context matters too: this Wan release cluster arrived in the same week as Kimi K3 (2.6T parameters, open weights), FLUX 3 (unified multimodal architecture announced by Black Forest Labs), and a wave of other open-weight releases — @aisearchio counted Wan-Dancer, Wan-Streamer v0.3, and Kimi K3 among eight major model releases in a single week. The open-weight ecosystem is accelerating across the board, not just within the Wan ecosystem.
For ongoing tracking of Wan 3.0 rumors, papers, and community analysis, the independent tracking page at Wan 3.0: Everything We Know aggregates updates as new information surfaces.
5. What Practitioners Should Do Now for Wan 3.0 Readiness
Deploy Wan 2.2 as Your Baseline
Wan 2.2 is stable, Apache 2.0, and has the largest community. The Diffusers API and safetensors format are consistent across Wan versions.
pip install diffusers transformers accelerate torch
Top HuggingFace downloads for reference:
| Model | 30-Day Downloads |
|---|---|
| Wan2.2-TI2V-5B-Diffusers | 166K |
| Wan2.2-I2V-A14B-Diffusers | 130K |
| Wan2.2-T2V-A14B-Diffusers | 109K |
| Wan2.2-S2V-14B | 74.2K |
Experiment with Wan-Dancer
It's the most recent Apache 2.0 release and demonstrates the hierarchical generation paradigm that a Wan 3.0 model would likely use. The Diffusers API is straightforward (see Section 2.2). Budget at least 32 GB VRAM for 14B bf16 inference — at 85 GB disk size (@wildmindai noted the download is substantial), plan for significant storage and bandwidth for the checkpoint alone.
Read the Papers
The architectural decisions in these papers will inform how a Wan 3.0 model works:
- WanSong — hybrid-MMDiT architecture, dual-stem output, RLHF alignment
- Wan-Dancer — hierarchical two-stage generation, global keyframe planning
- Wan-Streamer — world-event decomposition as pretraining objective
Integrate Audio Pipelines Now
Wan 3.0 will almost certainly output synchronized audio with video. The systems that adapt fastest when it ships will have the largest advantage in the creator tooling space. Every release points toward audio + video under a shared architecture.
Provision GPU Capacity for 30B+
If Wan 3.0 is a unified multimodal model, inference will require substantial GPU resources. WanSong is 25B for audio alone; a unified audio-visual model will need more. Budget for 4–8 A100 or H100 GPUs for the full checkpoint.
6. FAQ
Q: Has Wan 3.0 been officially announced?
No. Alibaba has not published a release date, model card, or technical report for Wan 3.0. The evidence in this article is based on three adjacent releases that share a multimodal backbone architecturally distinct from Wan 2.x. A Q3–Q4 2026 announcement is a reasonable working hypothesis, not a confirmed date.
Q: Is Wan 3.0 just Wan 2.2 with new features?
No. Wan 2.2 is a video-only DiT. WanSong, Wan-Dancer, and Wan-Streamer each use multimodal backbones that jointly process text, audio, and video tokens. A Wan 3.0 built on this architecture would be fundamentally different from Wan 2.2.
Q: Are WanSong, Wan-Dancer, and Wan-Streamer built on the same weights?
Unlikely. WanSong (~25B) and Wan-Dancer (14B) operate at different parameter scales with different architectures. Wan-Streamer's model size is undisclosed. They share architectural principles but are almost certainly independent checkpoints developed for different research goals.
Q: Will Wan 3.0 be open source?
Wan 2.1, Wan 2.2, and Wan-Dancer are all Apache 2.0. The trend is toward open weights, but WanSong and Wan-Streamer have no published license. Plan for either Apache 2.0 or a more restrictive option.
Q: Should I wait for Wan 3.0 before building on Wan?
No. Deploy Wan 2.2 now. It's stable, Apache 2.0, and well-documented. Experiment with Wan-Dancer to understand hierarchical generation. When Wan 3.0 ships, the migration will be smoother if you already have a Wan-based pipeline using the Diffusers API and safetensors format.
Q: How does Wan 3.0 compare to FLUX?
They target different markets. FLUX is a commercial multimodal model with API access and planned open-weight backbones. Wan, based on its Apache 2.0 trajectory, is likely to follow a fully open-source path. In the open-source video generation space, Wan 2.x already leads in HuggingFace downloads — Wan 3.0 would extend that lead across audio and real-time modalities.
7. Summary
The evidence for Wan 3.0's direction is visible across three July 2026 releases:
- WanSong proves a ~25B multimodal DiT can generate 5-minute songs with dual-stem output
- Wan-Dancer proves hierarchical two-stage generation works for minute-scale coherent video
- Wan-Streamer proves the world-event pretraining paradigm supports real-time streaming interaction
- All three share a hybrid-MMDiT backbone architecturally distinct from Wan 2.2's video-only DiT
- All three were released or published within one week
The playbook is straightforward: deploy Wan 2.2 as your baseline, experiment with Wan-Dancer for hierarchical generation, read the papers for architectural context, and provision GPU capacity for a 30B+ multimodal model. The Wan ecosystem is moving faster than any other open-source video project — the teams who prepare their pipelines now will be shipping features when Wan 3.0 arrives, not reading papers.
Browse Wan-AI on HuggingFace · Wan-Dancer model · WanSong paper · Wan-Streamer paper · Wan-Dancer paper