Sulphur 2 vs HappyHorse 1.0: The Benchmark Leader You Cannot Use Yet

For most "Sulphur 2 vs X" pages on this site, the comparison runs on capability and pricing. This one runs on a different axis: availability. HappyHorse 1.0 currently leads Artificial Analysis's video leaderboard across both text-to-video and image-to-video (no-audio category) with Elo scores noticeably above every closed commercial model — including Kling 3.0, Runway Gen-4.5, and Veo 3.1. That leaderboard position has driven a wave of attention and coverage in May 2026.

What is not currently true is that creators outside of the development team can use HappyHorse 1.0 to generate clips. The GitHub repository and the Hugging Face model card both display "coming soon" rather than downloadable weights. The project page on happyhorsemodel.ai describes upcoming public release plans but does not provide a generation endpoint a creator can sign up for today.

This creates a specific search pattern: people read about HappyHorse on a directory page, search for how to try it, and discover they cannot. Those searches are increasingly landing on alternative-pattern pages — "HappyHorse 1.0 alternative," "AI video model like HappyHorse," "open-source video model to try now." Sulphur 2 happens to be a credible answer to those queries because it is also an open-source video model from the May 2026 wave, also targets realistic output, and is actually usable right now.

The comparison below treats HappyHorse 1.0 fairly — it is a real and important model — but does not pretend creators can use it before access opens.

Sulphur 2 is an open-source AI video generation model distributed by the SulphurAI community on Hugging Face as SulphurAI/Sulphur-2-base. Released on 2026-05-03, it is a 9-billion-parameter fine-tune of Lightricks' LTX 2.3 (the 22B open-source base model), additionally trained on roughly 125,000 video clips with training data filtered to exclude 2D and animation content. The release ships with distill LoRAs, four ComfyUI workflows (T2V and I2V × base and distilled), and a local prompt enhancer. Running it locally needs a 24-32 GB VRAM workstation; community GGUF re-quants reduce VRAM needs to roughly 10-23 GB. sulphur2.net is an independent online hosting service that runs the same model behind a browser interface with credit-based generation.

HappyHorse 1.0 is a 15-billion-parameter video generation model from Alibaba's Taotian Group (Future Life Lab), led by Zhang Di — formerly Vice President at Kuaishou and technical lead on Kling AI. The model is described as a unified single-stream Transformer that jointly models text, image, video, and audio within one sequence, using a 40-layer self-attention architecture. Public materials list 5-to-8 second 1080P clips as the target output length, with reported lip-sync accuracy above 90 percent in English and Chinese. Authorship was confirmed by Alibaba on 2026-04-10. As of the time of writing, the project's GitHub and Hugging Face hub display "coming soon" rather than downloadable assets, and there is no announced standard API or subscription product.

The two models sit in the same broad category — open-weight video generation in 2026 — but are not currently usable in the same way.

HappyHorse 1.0's leaderboard position is the single most-cited fact in the model's public coverage, so it is worth being precise about what the numbers say and what they do not.

On Artificial Analysis, video models are compared through blind pairwise voting — a user sees two clips generated from the same prompt without knowing which model produced which, and picks the winner. Elo scores update from those votes continuously, the same mechanism used in chess rankings. HappyHorse 1.0 currently sits at Elo 1361 for text-to-video and Elo 1398 for image-to-video in the no-audio category. For comparison context on the same board: Kling 3.0 is at Elo 1243 (the highest among publicly available models), Runway Gen-4.5 at 1225, Veo 3.1 at 1217, and LTX 2.3 Fast — the base model Sulphur 2 fine-tunes — at Elo 1121.

What this means: when blind voters compare a HappyHorse 1.0 output against a clip from any of the other major models for the same prompt, HappyHorse wins more often. That is the real signal. It does not say anything about specific prompt categories (a model can win overall and still lose on a particular use case), and it does not say anything about audio behavior since HappyHorse's audio mode is not yet voting-tested at scale.

What this does not mean: it does not mean you can use HappyHorse 1.0 to generate a clip today. The leaderboard data comes from controlled access by Artificial Analysis to the model during testing; public access has not opened. Quoting the Elo number in a marketing claim ("we use the #1 model") would be misleading without the availability context.

Sulphur 2 has no Sulphur-specific Elo score yet — the model is barely three weeks old, and Artificial Analysis has not assigned it an independent ranking at the time of writing. The closest proxy is the LTX 2.3 base score (≈1121), which is the leader in the open-source-only sub-category. As Sulphur 2 accumulates community comparisons through May and June 2026, expect a separate Sulphur-2-specific entry to appear.

Sulphur 2 has a clear access story: open weights on Hugging Face, hosted browser access on sulphur2.net, 50 free credits at signup, credits do not expire, 6-month library retention. Pricing is on the pricing page. The model is downloadable and runnable on a 24-32 GB VRAM workstation today.

HappyHorse 1.0 has no public pricing because there is no public product yet. The project site at happyhorsemodel.ai describes the model's positioning and links to placeholder GitHub and Hugging Face pages that show "coming soon." There is no signup, no trial, no API key flow. Some community-aggregator pages list HappyHorse on their model catalogs, but these listings appear to be anticipating release rather than reporting current availability — verify the actual access status on the linked project page before planning around it.

For a creator deciding which to plan around: Sulphur 2 is the only one that supports planning around today. When HappyHorse opens access, this page will be updated within a week with the new specs, pricing, and availability path. Until then, the recommendation is to use Sulphur 2 for whatever short realistic clips your project needs and to bookmark this page for the HappyHorse update.

For Sulphur 2: the model is barely three weeks old and has no third-party benchmark yet. What can be said responsibly comes from inheriting the LTX 2.3 base behavior plus the visible effects of the realism-focused fine-tune. The LTX 2.3 base leads Artificial Analysis's open-weight category at Elo 1121. Sulphur 2's added training (~125K clips, 2D and animation explicitly filtered) biases it toward realistic faces, skin texture, atmospheric continuity, and microexpressions. Output quality on short realistic shots should be in the LTX 2.3 ballpark, possibly slightly sharper for realistic faces. For animation / 2D / illustrated prompts, expect weaker results than the LTX 2.3 base would have produced.

For HappyHorse 1.0: the model has Artificial Analysis Elo scores (T2V 1361, I2V 1398) but has not been used in production by independent creators yet. Public materials describe a 15B unified transformer with native audio-video generation and lip-sync above 90% in English and Chinese, but these claims have not been independently verified at scale because access is restricted. The Elo numbers are reliable as a directional signal — voters genuinely preferred HappyHorse outputs in blind tests during Artificial Analysis's evaluation — but quality on a specific prompt category cannot be confirmed without public testing.

What this means side-by-side: HappyHorse appears to be a stronger model overall once it opens, especially for any project that needs native audio or lip-sync. Sulphur 2 is a usable model today with a realistic-scene bias that fits a large set of creator use cases. The honest verdict is that both are interesting open-weight directions in the May 2026 wave, and a creator can use Sulphur 2 immediately while watching for HappyHorse's release.

The honest scenario list is short on this page because the choice is constrained by availability. For these scenarios, the 50 free credits are the right next move.

If you fit one of the scenarios below, the practical approach is: bookmark this page (we will update on release), watch the project's GitHub and Hugging Face placeholders for status changes, and run Sulphur 2 in the meantime for any silent visual work the project needs to start now.

"HappyHorse 1.0 is already the best model, so it is the right choice." Best on benchmark, yes — for now and for the prompts evaluated. But "best" only matters if you can use it. As of the time of writing, no public creator can. Picking a model that is not accessible is not a choice; it is a wait.

"Sulphur 2 is just a smaller HappyHorse." Both are open-weight video models from the May 2026 wave with realism-leaning training, but they are independent projects with different lineages (Sulphur 2 = LTX 2.3 fine-tune, 9B; HappyHorse = original 15B unified transformer with audio). Treating them as interchangeable understates what is different.

"Once HappyHorse opens, Sulphur 2 will be irrelevant." Possibly true on raw benchmark, possibly not on practical workflow. A 15B model with native audio will demand more VRAM than a 9B silent model. Sulphur 2 may stay the simpler tool for short silent clips even when HappyHorse opens; the question is which tradeoff fits the project.

"The leaderboard is the only thing that matters." Elo scores compress a lot of nuance. A model that wins overall can still lose on a specific use case. Use the leaderboard as a directional signal, not as the final answer for every project.

If the wait for HappyHorse is too long and Sulphur 2 does not fit, two other 2026 entries are worth considering. Each comparison page has the spec-by-spec breakdown.

HappyHorse 1.0 is the most interesting open-weight video model release of 2026 on paper — 15B parameters, native audio, ex-Kling leadership, and Artificial Analysis Elo scores above every other publicly evaluated model. The catch is timing: as of May 2026, weights and API are listed as "coming soon," not as downloadable or callable. Creators reading about HappyHorse cannot use it today.

Sulphur 2 is the closest available open-weight alternative in the same wave. It is smaller (9B versus HappyHorse's 15B), narrower (silent video versus HappyHorse's planned native audio), and not yet third-party benchmarked — but it is available right now, on sulphur2.net or locally, with 50 free signup credits.

For any silent realistic video project that needs to ship in May or June 2026 — use Sulphur 2. Start with the 50 free credits.

For projects that absolutely need native audio or lip-sync and can defer — bookmark this page (we update on release), use Sulphur 2 for the visual direction work in the meantime.

For research interest in top-end open models — track HappyHorse's GitHub and Hugging Face for the status change, and use Sulphur 2 today to build familiarity with the open-weight video workflow before HappyHorse drops.

Read the full Sulphur 2 Review for the deeper take, or the Sulphur 2 Showcase for prompt-ready examples grouped by use case.