fal's H3 Max: Post-trained MiniMax H3 Tops Image-to-Video Leaderboards

On Aug 27, fal released H3 Max, a video generation model post-trained from the open-weight MiniMax H3 and co-designed with fal's own inference stack, offered via the fal.ai API with MiniMax co-promoting the launch. It debuted at #1 in image-to-video (with audio) on Artificial Analysis' video leaderboard, narrowly beating ByteDance's Dreamina Seedance, and #3 in text-to-video; @isidentical noted it is the highest-ranked open-weight model on the Film Arena leaderboard, #4 overall out of 33 models, calling it an open-source video "DeepSeek moment"—though per @QuixiAI the model ultimately ships closed via API only, qualifying that framing.

Confirmed

• H3 Max ranks #1 in image-to-video and #3 in text-to-video on Artificial Analysis' leaderboard, with users able to vote in Video Arena; #4 overall on Film Arena, the highest-ranked open-weight model
• Post-trained for prompt adherence and aesthetics, with native audio generation, custom durations, and a tailored inference stack (examples include 768-resolution post-apocalyptic sci-fi scenes)
• fal claims #1 in overall quality, prompt understanding, and aesthetics across first-party and third-party evaluations (Artificial Analysis, Design Arena)
• Speed is the standout: official figures cite 5s 720p video in under 3 seconds with 35x the throughput of the official H3 endpoint; @isidentical measured 6.4s image-to-video and 4.7s text-to-video, 18x and 24x faster than the average model; @Recoil42 and @isidentical both report a new quality-speed Pareto frontier, with speedup figures ranging from 18/24x to 50x across posts
• Creator tests corroborate: a Japanese creator generated 15s video in 15 seconds and argued the near-instant iterate loop matters more than visual quality; other users report 30s video within 1 minute; @isidentical relayed 15s video in 21 seconds
• The model is also live on Venice and is a default option on several generation platforms
• Pricing is 50% off for the first week (some posts say through Sept 1); fal and MiniMax frame the release as showcasing the H3 base model's potential and advancing open multimodal models
• Per @QuixiAI, H3 Max is closed and API-only despite being derived from open weights

Why it matters

• H3 Max demonstrates that post-training—rather than training from scratch—can rapidly match or surpass leaderboard leaders in video generation
• With generation times compressed by an order of magnitude at high preference quality, the near-real-time iteration loop reported by creators transforms trial-and-error workflows, with direct value for latency- and cost-sensitive video pipelines
• fal, an inference platform, moving into post-trained models is a representative case of the "platform + in-house model" strategy; the open-to-closed API shift, however, qualifies the "open-source DeepSeek moment" narrative

2026-08-27 ~ 2026-08-29 · 29 related posts

• fal releases H3 Max: generates 720p video in 3s, ranks #1 on quality — jfischoff · 2026-08-27
• [source] MiniMax H3 Max sets video generation speed record: 50x faster than base model — isidentical · 2026-08-27
• Test: H3 Max generates 30s video in 1 minute — techhalla · 2026-08-27
• [source] MiniMax H3 Max tops video leaderboards via fal's post-training — ArtificialAnlys · 2026-08-27
• MiniMax H3 Max tops video generation leaderboards, generating 720p clips in under 3 seconds — ArtificialAnlys · 2026-08-27
• MiniMax H3 Max tops Artificial Analysis Video Leaderboards — ArtificialAnlys · 2026-08-27
• Test: H3 Max generates 9s video in 6.75s — chrisfirst · 2026-08-27
• fal's post-trained MiniMax H3 Max debuts #1 in Image-to-Video on Artificial Analysis — jfischoff · 2026-08-27
• MiniMax H3 Max sets new Pareto Frontier for video gen, 50x faster. — Recoil42 · 2026-08-27
• MiniMax H3 Max text-to-video API launches on fal.ai — venturetwins · 2026-08-27
• Fal releases H3 Max, generating video 35x faster than official endpoint — adamho · 2026-08-27
• Open-source AI video hits its DeepSeek moment: MiniMax H3 Max ranks 4th worldwide — isidentical · 2026-08-27
• MiniMax H3 Max hits fal: 15-second video generated in just 15 seconds — Hailuo_AI · 2026-08-27
• MiniMax H3 Max Generates 15s Video in 21s, Fastest Frontier Model — isidentical · 2026-08-28
• [source] fal Launches H3 Max: 5-Second Video in Under 3s, ~35x MiniMax H3 Throughput — MiniMax_AI · 2026-08-28
• fal releases H3 Max model generating video in under 10 seconds — altryne · 2026-08-28
• Omni, Qwen, and other Flash models tested; video generated under 10s — iScienceLuvr · 2026-08-28
• H3 Max Benchmark: 15x Faster Video Generation Without Quality Loss — gorkem · 2026-08-29

11 near-duplicate retellings: gorkem · isidentical · gorkem · DavidmComfort · isidentical · QuixiAI · gorkem · gorkem · jfischoff · isidentical · JenniferHli