RunningHub Launches H3 RH Enhanced: Post-Trained on MiniMax H3 Base, Claimed at 0.18 Yuan per Second
According to a September 29, 2026 report by Xinzhiyuan, RunningHub launched H3 RH Enhanced, a video enhancement model deeply post-trained on the open MiniMax H3 base. It claims that across 2,000 test sets, multi-shot long-sequence narrative significantly outperforms Seedance 2.0, and the character-collapse rate in extreme long scenes drops by more than 60 percent. The report gives two cost figures: the body says 0.18 yuan per second, while the conclusion says 0.1 yuan per second, so the vendor's latest billing statement should be used to verify.
What changed
On September 29, 2026, Xinzhiyuan reported that RunningHub released H3 RH Enhanced, a video enhancement model deeply post-trained on the open MiniMax H3 base. The report says it evaluated 2,000 controlled samples across four categories, 500 each: continuous character close-ups, multi-scene transition camera work, multi-character interaction, and full narrative with more than 16 shots. It says that in continuous generation of 8 to 16 shots, it significantly leads Seedance 2.0 on character identity consistency, scene element retention, and instruction constraint retention, performs close to Seedance 2.5 overall, cuts the same-character collapse rate by more than 60 percent in extreme long scenes over 16 shots, and has averaged nearly 1 million seconds of video generated per day since launch. On the engineering side, the report describes five layers of modification at RunningHub: scene-targeted post-training, resolution-aware adaptation, few-step distillation, attention and cache co-optimization (SageAttention2, in-house Cache-DiT, torch.compile, TP2 plus Ulysses4 hybrid parallelism), and batch serving scheduling; the training pipeline goes from scene SFT to consistency optimization to few-step distillation. The report says the model is listed on MiniMax's official site for global B2B creator access, with limited-time free trials on the RunningHub platform. The cost figures are inconsistent within the report: two places in the body say "0.18 yuan per second," while the conclusion and summary say "0.1 yuan per second." This draft follows the original body and records it as 0.18 yuan per second, subject to the vendor's latest billing statement. Statements such as "H3 was open-sourced in late July 2026" and "Elo 1130 on the Artificial Analysis video editing leaderboard" are the source's own paraphrases.
What it means for developers
For developers, the value of this news is not in leaderboard comparisons but in the fact that it sells "open base plus post-training plus inference and serving optimization" as a complete deliverable. If the claimed multi-shot consistency improvement holds, the biggest impact is on the scrap-rate structure of long-shot volume businesses such as short dramas, e-commerce, and comic dramas; the share of rework and manual retouching hours would fall, and the cost basis per qualified finished video would change, not just single-frame image quality. The report says multi-GPU setups use TP2 plus Ulysses4 hybrid parallelism, aimed at PCIe-connected multi-GPU environments rather than NVLink. If that premise holds, the evaluation focus should be whether existing RTX-class servers meet its deployment requirements, rather than considering high-end interconnect machines first. The report also says the delivery form bundles post-trained model, compute, and services, and that parent company HaimaCloud spreads costs through cloud rendering and off-peak inference scheduling. That means teams should first judge whether they lack model quality or operations and scheduling capability, since the two require different integration paths. It is important to distinguish: the 2,000 test sets, the collapse-rate reduction, the daily generation volume, and the cost basis all come from the publisher and the report, not independent verification; the comparisons with Seedance 2.0 and Seedance 2.5 should likewise be treated as source claims. Under the report's terms, the 0.18 yuan per second finished-video cost depends on batch scale, qualified finished-video rate, and scheduling efficiency all holding at once; it is not the unit price of a single scattered task.
Integration and evaluation
Recommendation one (conditional analysis, regarding the reported scene-targeted post-training): if your shot scripts often exceed 8 shots, or there are product silhouettes, character identities, and prop continuity that must hold across shots, it is worth running a small-sample comparison with your own footage rather than directly trusting the 2,000-set conclusion; if the business is single-shot talking-head, static product images, or short-video beat syncing, this benefit path is largely inapplicable and migration cost may exceed the gain. The premise is that you can obtain controllable inputs with the same reference material and prompts as your current model. A measurable evaluation standard is the retention ratio of character identity and key props after shot 8, and the number of reruns per finished video, compared shot by shot against your current model on the same script. Recommendation two (conditional analysis, regarding the reported batch serving scheduling and per-second cost basis): if you plan to migrate from a closed-source API, calculate the comprehensive cost per qualified finished video rather than the per-second generation price; the premise is that task volume is large enough for tasks with similar input shapes and compatible sampling configurations to form batches; if you only have scattered long-tail demand, the scheduling layer's amortization effect is limited and its quoted cost basis may not reproduce. A measurable evaluation standard is the combined cost per qualified finished video over one continuous week across GPU time, manual retouching hours, and failed rerun rate, compared on the same basis as your existing bill. Recommendation three (suggestion, regarding the reported multi-GPU deployment premise): if considering private deployment, first use existing PCIe multi-GPU machines to run a feasibility check with the reported parallel configuration; the premise is that memory and GPU count meet its requirements, and a measurable evaluation standard is single-video generation time and batch throughput under the same shot script, compared with platform-hosted calls. Recommendation four (suggestion, regarding weights and licensing): do not infer direct commercial usability from "post-trained on an open base"; first confirm in writing with the provider whether weights are released, what license is used, and what the commercial and redistribution terms are. The evaluation standard is obtaining explicit license text and authorization scope. The above are suggestions or conditional analysis, not facts stated by the source.
Limits and open questions
The 2,000 test sets, the over-60-percent collapse-rate reduction, the daily generation volume, the per-second cost, and other figures in the report all come from the publisher and the report's terms; no third-party reproduction is seen, and its evaluation method and baseline setup cannot be verified from the source. H3 RH Enhanced does not currently appear in KimiSeek's model catalog; this version is not currently listed on KimiSeek. The reported public access on MiniMax's official site and the RunningHub platform refers to vendor-side channels and is unrelated to whether KimiSeek offers the model. The cost figures are inconsistent within the source: two places in the body say "0.18 yuan per second," while the conclusion says "0.1 yuan per second." This draft uses the body's 0.18 yuan per second; readers should verify with the vendor's latest billing statement. Seedance 2.0 and Seedance 2.5 are mentioned only as effect references, with no verifiable evaluation conditions given. The cost basis depends on batch scale, qualified finished-video rate, and scheduling efficiency all holding at once; it may not hold for small-scale or long-tail task shapes, and price and availability will adjust over time. The report does not state whether the model's own weights are publicly released or what license is used, so "post-trained on an open base" cannot be used to infer that its weights can be freely used or commercialized; license and commercial terms must be confirmed with the provider. Statements about parent company HaimaCloud's cloud rendering and official MiniMax recognition are also source claims and have not been independently verified.