Pika H3 Flex: An Open Video and Audio Model Built on MiniMax H3, in Six Steps

Short answer: Pika has released H3 Flex, an open-weight video-and-audio model that is really two LoRA adapters on top of MiniMax's H3. It generates picture and synchronized stereo sound in six denoising steps instead of the 20 to 50 a standard H3 run needs, and it supports up to 12 reference inputs. It was reported at seventh place on DesignArena's text-to-video leaderboard, the highest open-weight entry at announcement. The catch is that it inherits the MiniMax H3 Community License, and you still need the large H3 base model to run it.

Much of the detail here comes from ComfyUI's documentation and AlphaSignal's report on Pika's announcement. Pika's weights are on Hugging Face and its inference code on GitHub, per AlphaSignal. We did not read Pika's own launch post, so leaderboard and speed numbers below are reported by Pika and not independently verified.

TL;DR

What Pika actually built

Pika did not train a new video architecture. ComfyUI's page describes H3 Flex as "a post-trained version of MiniMax H3": a fine-tune on a curated multi-task corpus of video, audio and reference examples, then distillation of the 50-step generator to six steps, then refinement with reward feedback and preference optimization. Because it ships as adapters, "no new model architecture is involved."

AlphaSignal's summary of the release breaks the two adapters into five workflows:

The Ref2VA adapter is the interesting one for creators. References can carry identity, costume, object shape, environment, motion (a reference video), and voice (an audio sample), while the prompt carries the action and the shot. ComfyUI also lists multi-shot direction: hard cuts, camera choreography and dialogue described inside a single prompt.

Why six steps matters

Diffusion-style video generators produce a clip by repeatedly cleaning up noise. Each cleaning pass, a denoising evaluation, is expensive for video plus audio. Cutting the count from 50 to six cuts latency almost proportionally. AlphaSignal reports the exact schedule uses timestep boundaries of 1000, 750, 500, 250, 100, 40 and 0, with fresh noise injected at intermediate stages, and says Pika's matched comparison shows roughly 1.8 times faster generation than MiniMax H3Max. Pika also claims its hosted API can produce a 15-second clip with audio in as little as seven seconds.

Two practical warnings from the ComfyUI docs and the README as relayed by AlphaSignal:

• Use the published schedule. A generic six-step Euler sampler follows a different process, and the adapter was trained for Pika's schedule. ComfyUI's templates use lcm with manual sigmas and CFG 1.0.
• Do not stack acceleration LoRAs. The adapter is already distilled, and ComfyUI says adding a turbo or Lightning LoRA on top "breaks the output."

How to try it locally

ComfyUI ships three template workflows under Template Library, then Video: text to video (needs ComfyUI 0.43.2 or newer), reference to video, and first and last frame (both 0.30.0 or newer). Each loads the H3 base plus one adapter through the standard LoRA loader node. The text-to-video workflow lists these downloads: the H3 diffusion model, the pika_h3_flex_fl2va_v1_bf16 LoRA, a Qwen3-VL based text encoder, and the H3 video and audio VAEs. Keep adapter strength at 1.0.

If you prefer the command line, AlphaSignal quotes the repository's CLI, which takes a task, model revision, resolution, aspect ratio, duration and seed. Representative requests were tested on a single 80 GB H100, and there is no multi-GPU sharding for one clip. A CPU-offload flag lowers VRAM at the cost of RAM and transfer time. Output lengths follow a 17n+5 frame grid, so a nominal 5-second request yields 124 frames.

Local output is limited to 480p and 768p. The hosted API reaches 1080p, 2K and 4K through draft-to-video upscaling, which Pika says preserves the draft's motion and audio.

Audio is the quiet headline

Most open video models still treat sound as an afterthought. H3 Flex generates dialogue, effects and music in the same pass as the picture. Locally you get a mixed stereo track. The hosted API can also return five separate WAV stems (dialogue, score, SFX, foley and ambience) for an extra $0.001 per generated second, which lets an editor drop them into a nonlinear editor or digital audio workstation without trying to separate a mix afterwards.

Pricing: hosted versus local

AlphaSignal reports launch pricing on the Pika API of $0.01 per output-video second at 480p, rising to $0.08 per second at 4K, described as roughly half Pika's regular pricing. Image and audio references carry no input charge; reference video is billed at the generation rate for input and output seconds. At those endpoints, 15 billed seconds cost $0.15 at 480p or $1.20 at 4K before stems and reference video. Pika also announced 15 free daily generations for API Club members during the first two weeks.

Local use trades money for hardware. With about 215.5 GB of files and an 80 GB-class GPU in the tested setup, "free" weights are not casual. If you do not own an H100, the hosted route or a rented GPU is the realistic path.

How good is it? The leaderboard lead is narrow

DesignArena ranks models by head-to-head user preference. As reported, H3 Flex recorded Elo 1293, seventh overall and the highest open-weight entry when Pika announced the release. Ahead of it, AlphaSignal lists Gemini Omni 1.1 Flash at 1360 and Wan 3.0 at 1324. The base MiniMax H3 sits at 1277, so the adapters add a measured but modest gain on that board, mostly in speed rather than raw quality.

Treat that as a snapshot. Arena rankings move daily, and a gap of a few Elo points means little. If video quality matters for your project, run your own prompts. For the base model's own benchmark story, see our MiniMax H3 launch coverage and the H3 Max developer guide.

The license catch

ComfyUI's page says the adapters are distributed under the MiniMax H3 Community License Agreement and asks users to review "the territorial and commercial-use conditions" before using outputs commercially. Our earlier reporting on the base release found that the license excludes local deployment in the United States, the European Union, the United Kingdom and South Korea. We have not confirmed whether Pika added any terms of its own, so check the adapter repository's license file yourself. If you are in an excluded region, the hosted API may be the only compliant path, but confirm that with Pika.

Who should care

Three groups gain the most from a fast open adapter like this. Indie filmmakers and motion designers can iterate on shots in about a minute instead of waiting through a 50-step run, which changes how many takes they can afford. Tool builders get a published recipe (subject to the license above) for distilling a large base model into a few steps, plus a working reference for reference-conditioned audio. Researchers get a real distilled checkpoint to study against its base, including how preference optimization affects motion and sound.

If you only need a single polished clip, a hosted model may still be easier. H3 Flex shines when you want control, repeatability and your own pipeline.

A filmstrip with one frame turning green, illustrating fast open-weight video generation with Pika H3 Flex

Where it fits among open video models

H3 Flex joins a quickly growing open field. FastVideo's FastH3 Omni also speeds up H3 with an eight-step reference-to-video-and-audio pipeline. Lightricks' LTX-2.5 is another open-weight video-audio model with multishot support. NVIDIA has shown real-time H3-class video faster than playback. The pattern: the open community is converging on one base, MiniMax H3, and competing on distillation, sampling schedules and control.

That convergence has a downside. If the base license restricts regions, every derivative inherits the restriction, and an ecosystem built on a single base carries a single point of legal risk.

What to do this week

1. Decide your route. Local needs an 80 GB-class GPU and 215 GB of disk; hosted starts at pennies per second.
2. Check the license for your country and use case before building a product on it.
3. Start in ComfyUI with the template workflows, keep the published sigmas, and do not add other speed LoRAs.
4. Test references. Ref2VA is where the model differs most from plain text-to-video; try a voice sample and a motion clip together.
5. Run your own comparison against the base H3 and another open model on 10 of your own prompts before trusting an Elo number.

Related reading

• MiniMax H3: open video model locked out of the US and EU
• MiniMax H3 Max: API and local developer projects
• FastVideo FastH3 Omni: eight-step open reference-to-video
• Lightricks LTX-2.5 open-weight video and audio
• NVIDIA real-time H3 video faster than playback
• Endless AI TV stream built on H3 Max

Sources: ComfyUI documentation, AlphaSignal, MiniMax on X.

Specs, prices and rankings are as reported on October 29, 2026 and may change; confirm the license and pricing before you build.

Update — October 30, 2026: Comfy's $25,000 open-source build contest, which offers a MiniMax H3 bonus, is in a reported dispute with Higgsfield: claimed vs verified.