fal's H3 Max cuts video generation costs 10x while hitting 1.5-second output
Source: a16z | Published: 2026-09-17T17:33:45Z
H3 Max delivers a 35x speed boost and 10x cost reduction for AI video generation. Amazon MGM has already shipped an internal AI production tool called Nara on fal's infrastructure, turning Hollywood from near-zero usage a year ago into fal's fastest-growing market.
One Saturday afternoon, Rohan, an engineer at fal, opened Twitch on his laptop and started streaming — continuously generating video with H3 Max and broadcasting it live. He used prompt techniques to maintain narrative continuity, with each shot automatically extending into the next. The stream wasn't announced in advance, but it spread quickly.
At the same time, prominent creator Levels.io reached out to fal on Twitter, saying he'd already built a website and wanted to set up his own infinite streaming experience. Meanwhile, a separate internal team at fal was independently building a continuous generation system with memory, transitions, and context awareness. Three threads running in parallel, each unaware of the others — then all three hit Twitter simultaneously.
Batuhan Taskaya, fal's engineering lead, described this as a moment that "happens at the company every few months": the entire team drops what they're doing and converges on the same thing.
Where the 35x Speedup Comes From
H3 Max is fal's post-trained version of Minimax's open-source model H3. Minimax H3 is what they consider "a genuinely next-generation open-source video model" — reference image support, a familiar architecture, frontier-level quality. fal had already run similar experiments on Ideogram and Flux image models and had the infrastructure foundation in place, so when H3 arrived, they went all in.
The optimizations stack in layers. The diffusion model's inference steps were compressed from 50 to 20, but fewer steps hurt quality — so fal first brought quality back up through an RL training process, then layered inference optimizations on top. The result: output quality equal to or better than the original model, at dramatically higher speed. On top of that, every component in the video generation pipeline was optimized individually — the LLM for prompt expansion, the diffusion model itself, the VAE that decodes latents back to pixels, and the super-resolution component at the end.
Hardware utilization per component rose from the 30–40% MFU typical of video inference to 70–80% — a figure approaching the theoretical ceiling for the industry. With each optimization layer stacking on the last, the result is the roughly 35x speedup fal published at launch.
The migration from Hopper to Blackwell chips alone delivers roughly 2–3x inference speed gains, but costs rise by a similar factor — so Blackwell's main value is in pushing the "super-realtime" threshold lower, enabling things like a 5-second video generated in under 3 seconds.
H3 Max Turbo: A 5-Second Video in 1.5 Seconds
Batuhan said the team originally planned to call this version H3 Turbo, but felt the name undersold the achievement — the quality is better than the original H3, and "Turbo" implies a cut-down version. So they went with H3 Max.
A week after launch, the inference team shipped another version: at 97% quality, 2x faster and 2x cheaper. It's called H3 Max Turbo and is now publicly available. Generating a 5-second video takes 1.5 seconds.
Batuhan's read is that speed and cost are good enough now. Over the next two to three months, fal's focus will shift to quality and controllability — not chasing further latency reductions.
2-Minute Memory, 60-Minute Continuous Video
The version Rohan streamed on Twitch connected adjacent clips by taking the last frame of one segment as the first frame of the next — no real memory, just visual continuity.
fal's ML team was simultaneously building something different: genuine contextual memory. The model retains the past 2 minutes of video in compressed form while maintaining an evolving system prompt that tracks overall world state, character relationships, and scene structure. Within 2 minutes, raw frame compression handles continuity; from the 2-minute mark to 60 minutes, an evolving summary prompt takes over.
Computational complexity grows exponentially with context length, and the 2-minute window is the pragmatic engineering ceiling. But with prompt techniques, users can continuously steer the narrative — effectively extending it indefinitely. fal has capped the maximum session length at 60 minutes.
This version launched publicly as H3 Max Director. Users can issue real-time commands — "a woman walks into the office" — and the model responds immediately, with the same characters still present, the same office, and the ability to cut back to prior scenes.
fal also built a demo site called fal Live where viewers vote on what happens next — "a banana walks in" or "the moon walks in." Multiple channels run in parallel with different themes: some fully random, some locked to an '80s animation aesthetic.
Hollywood Is the Fastest-Growing Segment
Batuhan cited a specific example: Amazon MGM Studios released an internal tool called Nara, built primarily on fal's infrastructure.
A year ago, Hollywood usage on fal's platform was essentially zero. Now it's the fastest-growing segment. fal's generative media conference was dominated last year by consumer AI creators; this year it's being driven by production companies — including new studios spun out of the majors specifically for AI content, as well as the in-house technical teams of those studios themselves.
Professional workflows look very different from consumer use. A popular approach right now: render a low-resolution scene layout in Blender, feed that video as a reference input to H3 Max, and get an output that's close to 100% controllable. Last week someone connected GPT/Astra-class LLMs to Blender for automated set dressing, then handed the result to H3 Max for rendering — opening up a new pipeline.
Speed is the critical variable in this workflow: fast means you can explore multiple directions in parallel rather than waiting on one result before deciding the next step.
Controllability Is the Core Focus Next
Controllability features fal is building or has already shipped:
Batuhan's emphasis isn't on individual features but on the maturity of fal's post-training infrastructure — it can now "onboard any new model and stack any new capability." It's no longer one-off training runs but a reusable toolchain. The target isn't "80% reliable" — it's "99.9% reliable", the level where professional users can genuinely depend on it.
Generative Video Is Reaching "Token Market Fit"
Batuhan used a specific phrase to describe the current market state for generative video and AI coding agents: "token market fit" — whether individual users can efficiently consume large volumes of tokens, using the product continuously in production workflows.
His threshold: roughly $10,000 in monthly token spend per user. By his account, there's already a cohort of professional users sitting at their computers doing video generation every day and spending thousands of dollars. The demand is real and has always been constrained by cost and speed. H3 Max cuts costs to a tenth of what they were and pushes speed up by more than an order of magnitude — significantly raising the ceiling for this group.
"Everyone is waiting for that mass consumer moment in AI," Batuhan said. He didn't predict what form it would take — but Rohan's unannounced Twitch stream may already be a preview.