Realtime models radar: week of October 5, 2026
What shipped or moved in realtime video and world models this week, each with its source: Odyssey-3, Vivix W1, Reactor's ComfyUI nodes, TaoMate-H3, FastH3 V2 and LingBot-Map.
Six things worth a developer's attention in realtime video and world models this week. Every item links to its primary source. Figures a vendor published about its own product are labelled as vendor-reported; we have not reproduced them.
Summary
| Item | What it is | Source date |
|---|---|---|
| Odyssey-3 | A world model for generating and predicting environments in real time, in research preview | 8 October |
| Vivix W1 | A streaming video model with a published per-second price | Page checked 11 October |
| Reactor for ComfyUI | Node pack that runs Reactor's realtime video models inside ComfyUI workflows | Repository updated 10 October |
| TaoMate-H3 | A 3-step LoRA streaming runtime on MiniMax H3 from Alibaba's TaoLive team | Repository from 5 September; covered 9 October |
| FastH3 V2 and Trim | Consumer-GPU timings for the 8-step FastH3 release | 6 October |
| LingBot-Map | An open streaming 3D reconstruction model, ECCV 2026 best paper candidate | Repository updated 10 October |
Odyssey-3: a world model aimed at physical AI
On 8 October Odyssey announced Odyssey-3, described as a learned dynamical system implemented as an autoregressive diffusion transformer. According to Odyssey, it generates embodied environments from a prompt and predicts in real time how they change as a person or agent acts. A research preview is available, and the company is inviting physical-AI developers to get in touch.
Odyssey reports a new result on the Physics-IQ Verified benchmark for Odyssey-3 Pro, and a first-place ranking in three of four WorldMark categories in its own evaluations. Those are vendor-reported results.
Why it matters: world models are being positioned as simulators for training and testing robots and agents, not only as game-like generators. Source: Meet Odyssey-3, 8 October 2026.
Vivix W1: a streaming model with a per-second price
Vivix describes W1 as a streaming-native multimodal model for worlds that evolve in real time through text, voice and motion. Its product page lists $0.003 per second, labelled a limited-time discount through 31 October, so the price after that date is not published.
Vivix's own comparison, dated 8 October, lists other products by USD per output second: MiniMax H3 Max on fal at $0.080, Kling 3.0 (Omni Std) at $0.112, Seedance 2.0 at $0.151 and Seedance 2.5 at $0.231. It also reports an average 8.4 seconds to produce an 8-second reference-to-video clip at 720p, against 10.8 seconds for H3 Max on fal and 145.9 to 272.4 seconds for the others (average of 10 runs, queue time excluded).
Read the comparison with care. Those are vendor-reported figures; the comparison is per second of output clip, whereas a realtime service may meter session time. The billing start and stop conditions depend on the provider. On Windflow, billing starts at the first generation_started event; see billing boundaries. Output seconds and billable session-seconds are not the same unit. Source: Vivix W1.
Reactor for ComfyUI: realtime models as workflow nodes
Reactor published comfyui-reactor-api, a ComfyUI custom node pack for its realtime video models, under Apache-2.0. Its README says you can play a model live in a window you steer as it generates, or script a sequence of segments and render them to a video, from an ordinary ComfyUI workflow. The example shows LingBot World 2 steered with W, A, S, D and the arrow keys.
Why it matters: node packs put realtime models in front of the creators who already live in ComfyUI. Source: reactor-team/comfyui-reactor-api.
TaoMate-H3: a streaming runtime on MiniMax H3
Alibaba's TaoLive AIGC team published TaoMate-H3, a low-latency streaming audio-video generation runtime built on MiniMax H3 with a 3-step LoRA. It generates synchronized audio and video in small chunks and continues long-form generation at 480p, 768p and 1080p, in portrait or landscape. Inference is single-node on 4 or 8 GPUs with tensor parallelism and Ulysses sequence parallelism. The T2AV LoRA weights are on Hugging Face; the repository lists an FL2AV variant as still to be released.
The repository dates from 5 September and the first news entry from 8 September; Jiqizhixin covered it on 9 October, which is why it surfaced this week. TaoMate-H3 ships LoRA weights for the MiniMax H3 base model, which has its own licence, so read both the repository's terms and the base model's before building on it. Source: TaoLiveAIGC/TaoMate-H3.
Windflow serves MiniMax FastH3, a different distillation of the same base model, as a preview.
FastH3 V2 and Trim: timings on consumer GPUs
Hao AI Lab published end-to-end timings for the FastH3 V2 release, from prompt to MP4 for a 5-second clip with audio. Figures are for V2, with the Trim variant after the slash:
| GPU | V2 / Trim |
|---|---|
| RTX PRO 6000 | 13.5 s / 12.0 s |
| RTX 5090 | 14.8 s / 13.4 s |
| RTX 4090 | 54.6 s / 43.9 s |
| DGX Spark | 141 s / 126 s |
| 2 × DGX Spark | 87 s / 78 s |
At 768p, the thread reports 35 to 39 seconds on one RTX 5090 or RTX PRO 6000. These are vendor-reported timings of offline generation, not streaming latency. The weights are on Hugging Face as FastVideo/FastVideo-FastH3-8-Step-V2, published under a licence marked "other" on the model page. Source: Hao AI Lab thread, 6 October.
LingBot-Map: streaming 3D reconstruction
Robbyant, Ant Group's embodied-intelligence unit and the team behind LingBot-World, released LingBot-Map, a feed-forward 3D foundation model for streaming reconstruction under Apache-2.0. The README reports stable inference at about 20 FPS at 518 × 378 over sequences longer than 10,000 frames, and the repository is marked an ECCV 2026 best paper award candidate. It had about 18,000 GitHub stars on 11 October.
It reconstructs 3D from video rather than generating video, so it sits next to realtime world models instead of replacing them. Source: Robbyant/lingbot-map.
From Windflow this week
- Our ROCm fix for SANA's fused Gated DeltaNet kernels was merged upstream on 11 October as Sana #510, which unblocks SANA-WM and SANA-Streaming on AMD Instinct. The full list of fixes and measurements is in the public measurement report.
- Model pages now carry upstream sources, licences and comparison tables; start at the model catalog.