Context

a16z's Jennifer Li talks with fal cofounder Gorkem Yurtseven and head of engineering Batuhan Taskaya about H3 Max, fal's post-trained version of MiniMax's open-weight video model, and what changed once video generation got fast and cheap enough to run in real time. The conversation matters to PMs building on generative media because it's a concrete case study in how post-training plus systems optimization (not a new base model) unlocked entirely new product categories, live, continuous, steerable video streams, that didn't exist as usable experiences a few weeks earlier, and in how a team's own unplanned internal experiments became the company's most viral launches.

The Big Idea

Once a generative model gets fast and cheap enough (fal made video generation roughly 35x faster and an order of magnitude cheaper through post-training and systems co-design, without sacrificing quality), the constraint that matters shifts from speed to control, and control turns out to be the harder, more valuable problem for professional and Hollywood-grade use cases.

fal's own roadmap reflects this directly: after achieving speeds like generating a 5-second video in 1.5 seconds, the team explicitly pivoted its next one to two months of work toward camera control, lighting, lip sync, and character motion, concluding that "people already had the speed and cost they needed, the real gap left was fidelity to a director's specific intent."

Key Insights

Post-training an open-weight model, not building a new one, produced the breakthrough

fal didn't train a new video model from scratch; they took MiniMax's open-weight H3 model (chosen specifically because it was the first truly next-generation, capable open-weight video model available) and post-trained it, using reinforcement-learning-style pipelines to first raise quality at a reduced step count, then layered custom kernels and systems engineering on top. The compounding result: roughly 35x faster generation than the original MiniMax endpoint, at comparable or better quality (measured by ELO score), and an order of magnitude cheaper. The lesson for teams building on open-weight models: the model weights are a starting point, not a ceiling, real differentiation can come from post-training and inference-stack work layered on top rather than from training a bigger model.

Compounding small optimizations across pipeline stages beats optimizing one component

Taskaya describes video generation as a multi-stage pipeline (prompt expansion via a large language model, diffusion-based generation in latent space, decoding latents back into pixels, sometimes an upscaling step), each stage unoptimized by default and each requiring a different optimization approach (the prompt-expansion LLM behaves like a single-shot, low-batch workload very different from a caching-heavy coding-agent workload; the diffusion and decoding stages have their own distinct constraints). fal's team pushed hardware utilization from the industry-standard 30 to 40% up to 70 to 80% of theoretical maximum, essentially reaching the practical ceiling for the current hardware, and the 35x speedup came from stacking every one of these component-level gains together rather than from any single fix.

The team explicitly traded a "safe" name for an accurate one

fal initially planned to call their optimized model "H3 Turbo," signaling a faster-but-lesser variant, but reversed the decision once evaluations showed it matched or exceeded the original model's quality while being dramatically faster, renaming it H3 Max specifically so the name wouldn't undersell the achievement. They then held the public launch for three to four days specifically to get independent, external evaluation platforms to confirm their own internal (and expensive, tens of thousands of dollars in eval spend) results before announcing, because the internal numbers looked "too good to be true" and they wanted third-party validation before making the claim publicly.

Unplanned, parallel internal experiments produced the most viral launches

Once H3 Max shipped, employees across separate, independent teams (one engineer personally livestreaming continuous AI-generated video on Twitch, a formal "world model accelerator" team building a WebRTC live-streaming experience, an outside Twitter influencer building his own hosting site) each built and shipped their own experimental product around the same new capability within days, without central coordination. fal's cofounder frames this as a recognizable company-wide pattern: periods of quiet, incremental work punctuated by a capability threshold crossing that triggers spontaneous, parallel, fast-moving experimentation across the whole team, rather than a single planned launch.

Continuous video memory required exponentially harder optimization, not just longer clips

fal's early attempts at extending clip length by simply stitching independent 5-second clips together (using only the last frame as context for the next clip) produced visibly discontinuous scenes with no real memory. The eventual H3 Max Director model instead attends to roughly two minutes of actual generated video directly (computationally expensive, since attention cost scales up sharply with context length) and maintains a separately evolving, more compressed "system prompt" style summary for anything before that two-minute window, letting a scene maintain continuity (a character who left the frame is still there when the camera pans back) across streams up to an hour long.

Camera control required structured, explicit specification, not better prompting

fal's camera-control feature doesn't rely on natural-language description of camera movement; users supply a structured, timestamped specification (effectively a JSON schedule of camera position and angle at each point in time), and the model is tuned to treat that structured input as ground truth rather than a suggestion it might reinterpret. Yurtseven is explicit that this required actually tuning the model to respect the structured input reliably, prompting alone gets roughly 80 to 90% reliability, while fal's target for professional use is 99.9%, precise enough that a studio can trust the model executed a specific instruction exactly rather than approximately.

Hollywood's fastest-growing need is small, specific point solutions, not full generation

Yurtseven says Hollywood usage was close to nonexistent a year earlier and is now fal's fastest-growing segment, and the actual demand isn't "generate an entire scene from a prompt," it's narrow, composable capabilities: extend an existing shot slightly, change only the camera angle, adjust only the lighting, sync only the lip movement to new audio. He frames the gap explicitly: "what Hollywood needs and what creators actually need and what the research labs are working on" are currently disconnected, and fal's specific business is closing that gap with targeted post-training projects built directly from what working studios tell them they need, rather than from what a lab assumes a professional workflow requires.

Mental Models & Frameworks

Token market fit: can one person productively spend a huge token budget?

fal's framing for why generative media (alongside coding agents) is a uniquely large opportunity: the test isn't just "do people want this," it's whether a single professional user can productively spend on the order of $10,000 worth of tokens doing their actual daily job, generating and iterating on video for hours at a stretch. This reframes capacity planning: fal describes being compute-constrained industry-wide since roughly April, meaning the binding constraint on revenue and usage isn't demand, it's how much of that already-proven willingness-to-spend the available compute can actually serve, which is precisely what the speed and cost optimizations were built to relieve.

The controllability ladder: text, image, reference, then structured direction

fal traces generative video's evolution as a sequence of control primitives, each solving what the previous one couldn't: text-to-video (a prompt in, a video out, but no way to specify a particular character); image-to-video (an initial frame anchors the first moment, but new elements couldn't be introduced mid-scene reliably); reference-to-video (a starting frame plus named characters and voices, letting a specific character persist across a scene); and now structured, timestamped controls for camera position, lighting direction, lip sync, and motion. Use this ladder to diagnose where a video-generation product actually sits: a tool still limited to text-to-video is solving a fundamentally different, earlier-stage problem than one offering structured camera and lighting control, even if both are labeled "AI video generation."

Practical Application

Separate "faster" from "better" when evaluating a post-trained model variant

Before publishing or adopting a faster, cheaper model variant, explicitly check whether quality actually held (via a quality metric like ELO score) rather than assuming speed and cost gains come at a hidden quality cost. fal's own H3 Max Turbo variant runs twice as fast as H3 Max but does have a small, real quality drop, and the team surfaces that tradeoff explicitly rather than only advertising the speed number, so buyers can choose the right variant for their specific use case.

Hold a surprising internal result for independent validation before announcing it

When an internal benchmark result looks unusually strong (fal's early internal evals on H3 Max looked "too good to be true"), delay the public claim specifically until an independent third party can reproduce or confirm it, even if that costs days of delay on an otherwise-ready launch. This converts a claim that invites skepticism into one backed by external evidence at launch.

Give teams room to build spontaneously right after a capability unlock, rather than only shipping the planned roadmap item

When a genuine capability threshold is crossed internally (not an incremental improvement, but something that changes what's newly possible), consider deliberately not immediately constraining follow-on work to a single planned deliverable. fal's most viral products came from letting multiple teams and even outside collaborators experiment freely and in parallel right after the H3 Max launch, rather than funneling all engineering effort into one predetermined follow-up feature.

Ask professional users what narrow capability they need, not what full workflow they want automated

When building AI tools for a professional creative or technical audience (Hollywood, in fal's case), resist defaulting to "generate the whole output from a prompt." fal's direct engagement with studios revealed the actual demand was narrow, composable point solutions (only the camera angle, only the lip sync, only an extension of an existing shot), which is a materially different, more targeted product surface than a general-purpose generation tool.

Bottom Line

fal's H3 Max shows that once a generative model is fast and cheap enough to remove speed and cost as constraints, the real differentiator becomes precise, structured controllability, camera position, lighting, character consistency, specified explicitly rather than left to prompt interpretation, and that shift changes what professional users actually need from "generate this for me" to "execute this exact instruction reliably."