> ## Content Index > Fetch the complete content index at: https://www.tao.media/llms.txt > Use this file to discover other available public pages before exploring further. # Affine Launches MiniMax-H3 Video Studio and SN120 Sparse-Attention Competition - URL: https://www.tao.media/affine-launches-minimax-h3-video-studio-and-sn120-sparse-attention-competition/ - Published: 2026-10-21T02:19:36.000Z - Updated: 2026-10-21T02:19:36.000Z - Description: The Bittensor subnet is rewarding better attention selection to reduce video-generation compute while preserving output quality. - Author: Tristan Hillerich - Tags: Affine, Subnets, AI, News Affine has launched [Affine Video](https://video.affine.io/?ref=tao.media), pairing a public MiniMax-H3 video-generation studio with a new mining competition on Bittensor subnet 120\. The competition rewards improvements to the component that selects which attention blocks the video model computes, targeting lower inference costs without changing the underlying generator. Bittensor co-founder Jacob Steeves [announced Affine Video on X](https://x.com/const%5Freborn/status/2112567604955238635?ref=tao.media), calling it the "first model release" and sharing a generated clip of a ship at a foggy dock. The launch adds a video-focused mechanism alongside Affine's existing language-model training work. Its miners submit trained selector weights rather than operate a video-serving endpoint, and the winning selector receives all reward weight allocated to mechanism 1, not the whole subnet's weight. ## Inside the MiniMax-H3 Studio The [Studio](https://video.affine.io/studio?ref=tao.media) generates video with audio using the open-weights [MiniMax-H3](https://huggingface.co/MiniMaxAI/MiniMax-H3?ref=tao.media) model. Its interface offers nominal five-, 10-, and 15-second clips in landscape, portrait, 4:3, and square formats, priced at one, two, and three credits, respectively. The [live configuration](https://video.affine.io/api/config?ref=tao.media) sets a daily allowance of 20 credits and 24-frame-per-second output. Landscape clips use a 1344-by-768 resolution. The configured generation lengths are approximately 5.17, 10.13, and 14.38 seconds, corresponding to the interface's rounded duration labels. A script stage turns a user's idea into a generation prompt. Pro mode bypasses that step and sends a model-ready prompt directly to MiniMax-H3, checking its format as the user types. Users can also set a seed. Although the underlying MiniMax-H3 model supports first- and last-frame conditioning and reference-based generation, the Studio's current configuration enables text-to-video and disables reference uploads. The Studio is configured to run eight denoising steps on a single RTX 5090 with 32GB of memory. Its estimates for landscape generation range from roughly 80 seconds for the shortest clip to about four minutes for the longest. The backend was offline at the time of reporting. Earlier, Steeves wrote that "[our servers immediately got destroyed](https://x.com/const%5Freborn/status/2112583574863937602?ref=tao.media)" after sharing the launch, attributing the disruption to demand. ## Training the Selector, Not the Generator Video diffusion models build clips through successive denoising steps. Attention links information across tokens representing the scene, helping preserve spatial relationships and motion over time. Full attention computes connections across the entire sequence, making it expensive as video sequences grow. Sparse attention skips selected blocks of those connections. The selector must identify useful blocks without discarding information the generator needs for coherent output. Affine miners train the selector's query and key projections while the base video model, inference code, and eight-step schedule remain fixed. The approach builds on [Veda, an ICML 2026 sparse-attention method](https://veda-sparse.github.io/?ref=tao.media) from researchers at ByteDance, the University of Hong Kong, and the University of Science and Technology of China. Veda distills attention patterns from a full-attention teacher into a lightweight predictor that selects blocks for a sparse execution kernel. The researchers report quality comparable to full attention at 90% sparsity. In a separate 95%-sparse configuration, they report generating a 720p, 241-frame Waver-T2V-12B video in 3.8 minutes instead of 19.4 minutes, a 5.1-fold end-to-end speedup. Those are Waver research results, not measured gains from Affine's MiniMax-H3 competition. The [MiniMax-H3 reference selector](https://huggingface.co/Veda-Sparse/Minimax-H3-T2VA-Veda-8NFE-600Step-Preview?ref=tao.media) supplies the starting incumbent. Miners choose their own training methods and datasets; copying its weights does not establish an improvement. ## How the SN120 Competition Works Under the [VEDA submission rules](https://github.com/angosr/affine-video?ref=tao.media), miners upload `selector.safetensors` to a public Hugging Face model repository, pin a full 40-character commit SHA, and commit that reference on Bittensor's Finney network. Miners do not need to run a server or expose a public inference endpoint. The file contains exactly 100 float16 tensors: query and key projections for each of 50 layers. Each tensor has shape `(56, 384, 128)`, and the file may not exceed 560 million bytes. Metadata binds the artifact to the miner's hotkey, its operational signing identity. Base-model weights, adapters, and executable code are excluded; the evaluator never runs code from the submission repository. Each eligible hotkey gets one submission opportunity on the track. A malformed artifact, an invalid model, or a lost challenge consumes it. Hotkeys activated in Affine's original competition are ineligible for VEDA. Candidates enter a first-in, first-out queue and face two evaluation stages. First, challenger and champion run on the same hidden batch of 16 videos across four configurations. Eight sampled layer/head routes and three denoising steps produce 384 paired cases, evaluated at attention-block budgets of 16, 32, and 64. The [track specification](https://iblosevwkuiootbbgnsy.storage.supabase.co/storage/v1/object/public/affine%5Fvideo/tracks/main/track.json?ref=tao.media) sets a minimum frontier gain of 0.005 with 99.75% confidence per comparison. A 10,000-resample bootstrap groups correlated cases by video. Scoring combines attention quality with calibrated compute cost. Candidates that pass then undergo full eight-step video generation at all three budgets on a separate hidden batch. That stage combines latent, frame, and temporal fidelity with weights of 0.4, 0.3, and 0.3 and requires a minimum gain of 0.005\. Only a candidate clearing both stages becomes champion. Ties retain the incumbent; infrastructure failures trigger retries rather than miner losses. ## The Reference Selector Holds the First Match The [public leaderboard](https://iblosevwkuiootbbgnsy.storage.supabase.co/storage/v1/object/public/affine%5Fvideo/tracks/main/status.json?ref=tao.media) lists track `veda-sn120-v2-r4` as accepting submissions. Its [match summary](https://iblosevwkuiootbbgnsy.storage.supabase.co/storage/v1/object/public/affine%5Fvideo/tracks/main/summary.json?ref=tao.media) recorded one completed challenge and two leaderboard entries at the time of reporting. In the [first match](https://iblosevwkuiootbbgnsy.storage.supabase.co/storage/v1/object/public/affine%5Fvideo/tracks/main/matches/veda-sn120-v2-r4/000000000001.json?ref=tao.media), UID 472 scored 0.94241 against the reference champion's 0.94239\. Its paired-gain lower confidence bound was approximately −0.0000376, below the required 0.005\. The challenger failed the first stage and never reached full-video confirmation, leaving UID 0 in place with a `CHAMPION_RETAINED` outcome. These are quality/cost scores, not inference-speed multipliers. The [published weights file](https://iblosevwkuiootbbgnsy.storage.supabase.co/storage/v1/object/public/affine%5Fvideo/tracks/main/weights.json?ref=tao.media) assigns UID 0 a weight of 1.0 within mechanism 1\. Validators use that file to set mechanism-specific weights, and the champion retains the allocation until replaced or no longer eligible. Affine's separate [MiMo GRPO model-training run](https://affine.io/llms.txt?ref=tao.media) went live on subnet 120 on October 13\. VEDA participants can follow their own track through the public board or the read-only `veda monitor` command, which needs neither a wallet nor a GPU.