Pika H3 Flex: Open Weights, 6-Step Speed, and How to Run It

On October 29, 2026, Pika released H3 Flex, its new flagship video model — and did the thing few flagship labs do anymore: it published the weights. H3 Flex generates video with synchronized stereo audio built in, samples in 6 denoising steps, and sits on Hugging Face for anyone to download. Three days later, on November 2, it rolled into Pika's creation product for people who never want to touch a checkpoint file.

The name is a family tree: H3 Flex is Pika's post-trained version of H3, the open video model from MiniMax — the H3 in the name is the base model. Pika frames the result as an "omni-generator": one model that outputs picture and sound together, steered through references rather than prompt incantations.

This guide covers what the weights actually include (and what stays hosted), how to run it locally in ComfyUI, the API path with its reported pricing, and the license caveat that should slow you down before any commercial use.

What H3 Flex is, and where the H3 name comes from

Pika took MiniMax's open H3 video model and post-trained it. The result ships with two task adapters: FL2VA, which conditions on text plus an endpoint frame (first/last-frame video), and Ref2VA, which takes up to 12 mixed reference files — images, video, or audio — and generates from them. References also drive editing: swap an outfit, a prop or a setting while keeping the performance intact.

Distillation is the second headline. The teacher model ran roughly 50 denoising steps; H3 Flex needs 6 — about 8x fewer. At launch, Pika reported #1 among open-weights models and #7 overall on DesignArena's text-to-video leaderboard. That's Pika's claim about a launch snapshot, not an independent benchmark we've reproduced — but it's specific enough to check, and consistent with how fast the community picked the model up.

The two adapters: FL2VA and Ref2VA

The adapter names decode the input contract:

In practice, most briefs are Ref2VA briefs: hand the model the subject as files, use the text to describe the shot. It's the same discipline our prompt guides preach — fill every slot with something concrete — except here, three of the slots can be files.

The specs in one table:

What "open weights" actually gets you

Downloading the checkpoint is not the same as getting the product. The honest split:

The core generation loop — audio included — runs anywhere you can fit the model. Everything that depends on infrastructure Pika controls (the upsampler, the stem splitter, prompt rewriting, their optimized serving stack) lives behind the product and the API. That's the standard open-weights bargain; H3 Flex just draws the line unusually clearly.

The split also explains the pricing gap discussed around the model. A self-hosted H3 Flex render bills in GPU-hours; the hosted render bills per second of video, but reaches 4K and hands you stems. Neither is free — one charges in hardware, the other in metered fees.

Running H3 Flex locally

The ecosystem moved within days of release:

• ComfyUI has native support from v147, and SwarmUI works too.
• Pika's own Turbo-LoRA ships 2-, 4- and 8-step checkpoints and became the most-downloaded H3 Flex variant within a week — roughly 43,000 downloads by November 5.
• Community artifacts: WAS/Pika-H3-Flex-PTD, a FastVideo NVFP4 quantization, and an MLX port for Apple Silicon. The MLX port is community work, not something Pika maintains.

A minimal local path

1. Update ComfyUI to v147 or newer.
2. Download the model from the Hugging Face repo (it ships in diffusers format).
3. Validate your setup with the base 6-step workflow at 480p.
4. Once it works, swap in the Turbo-LoRA checkpoint and pick a 2-, 4- or 8-step variant.
5. Budget for hardware before anything else: without H100-class VRAM, you'll feel it immediately.

Hardware is the honest caveat. The full model is H100-class GPU territory for comfortable use. The Turbo-LoRA cuts sampling steps, not the size of the base weights, so it speeds up generation without making a consumer GPU suddenly sufficient. Without datacenter silicon, the local route is "possible, expensive, and mostly for people who already run inference servers" — though Mac owners should watch the MLX port, with the usual community-port caveats about quality and memory.

The API path

For everyone else: Pika's product integrated H3 Flex on November 2, and API access is sold through what Pika calls the Pika API Club. Third-party reporting puts pricing around $0.01 per second of video — we couldn't find a published rate sheet on Pika's own pages at recording time, so treat that number as reported, not confirmed.

What going hosted buys you: the upsampler (1080p to 4K), five-stem audio separation, prompt enhancement, and accelerated serving. For production work where stems matter — dialogue you need to mix, score you need to replace — the API is the only route to separated tracks today.

Before committing budget, three checks: confirm current pricing on Pika's official API page (the $0.01/sec figure is reporting, not a rate card), verify which resolutions and stems your access tier actually includes, and validate prompts at 480p rather than paying for 1080p upsampling on every iteration.

How H3 Flex stacks up

We won't print a head-to-head scoreboard: leaderboards shift weekly, and we only publish numbers we can verify. What's stable is the shape of the trade-off, on three axes:

• Open weights. The closed flagships — Sora and its peers — are product-only. H3 Flex is a rare top-tier option whose full checkpoint you can download, inspect and build on. That matters little for a one-off clip and a lot for anyone with a pipeline, a fine-tuning plan, or a data-sensitive workload.
• Native audio. H3 Flex generates synchronized stereo sound alongside the picture. Most video models still treat audio as a separate post step, and native synced audio is rare among open-weights models at this tier. Pika's own leaderboard framing carried the same qualifier: #1 among open-weights, #7 overall against closed peers.
• Speed. Six steps instead of ~50 is a real architectural difference, not marketing. In Pika's controlled benchmark, a 15-second 768p clip took as little as seven seconds of transformer (DiT) compute — the denoising work itself. That is not the end-to-end time you'll wait for an API response, and Pika says as much: queueing, encoding and delivery each add their own seconds.

Against open peers, the fair comparison is capability breadth: reference-to-video with audio, first/last-frame control, and reference-guided editing in one checkpoint is a lot of surface area for an open release. Whether it beats a specific rival on your footage is an empirical question — run the same prompt through both and look.

Who it's for

• Run it locally if you're an engineering team, a research group, or a studio that wants the weights: fine-tuning, batch pipelines, air-gapped or data-sensitive work. Bring H100-class hardware and read the license first.
• Use the product or API if you want results: 4K output, separated stems, prompt enhancement and fast serving are hosted-only anyway.
• Pause if you need guaranteed commercial terms without reading legal text, or if your core workload is photoreal dialogue scenes at feature-film fidelity — no open model is a safe default there yet.

Read the license before you ship anything

Hugging Face lists H3 Flex's license as "other" — under the model's own license terms, specifically the MiniMax H3 Community License Agreement that Pika's release inherits from the MiniMax base. That is not a standard open-source license, and "community license" names have covered everything from extra conditions to revenue thresholds to field-of-use restrictions. Before any commercial deployment, read the actual LICENSE file in the model repo — the Turbo-LoRA repo carries the same terms. If the terms don't work for you, the hosted API sidesteps the weights license entirely, under Pika's commercial terms instead.

Where to go next

The prompting habits that carry across video models — structure over magic words, one camera move per shot — apply to H3 Flex too, especially when briefing Ref2VA references; see our Jimeng AI prompts guide and Seedance 2.0 prompts for the formula. To compare hosted rivals, look at Runway and Kling AI, or browse the rest of our AI video coverage. Primary sources: Pika's launch post, the research deep-dive, and the weights on Hugging Face.

FAQ

Is Pika H3 Flex open source?

It has open weights — the checkpoint is downloadable from Hugging Face and GitHub — but that is not the same as open source. The license is custom (the MiniMax H3 Community License Agreement), so commercial use depends on what that text allows, not on open-source norms. Read it before you build a business on it.

Can I run Pika H3 Flex on my own hardware?

Yes, with realistic expectations: the full model wants H100-class GPUs. Pika's Turbo-LoRA (2/4/8-step) speeds up sampling and runs in ComfyUI v147+, and a community MLX port targets Apple Silicon — but fewer steps shrink generation time, not the model's memory footprint.

How much does the Pika H3 Flex API cost?

Third-party reports put it around $0.01 per second of video, sold through Pika's "Pika API Club". We couldn't verify a published rate sheet, so confirm current pricing on Pika's official page before budgeting.

What's the difference between Pika H3 Flex and MiniMax H3?

MiniMax H3 is the open base model; H3 Flex is Pika's post-trained version on top. Pika's contributions: distillation to 6 denoising steps (~8x fewer than the ~50-step teacher), the FL2VA and Ref2VA task adapters, and the reference-and-editing workflow with synchronized audio.

Does H3 Flex generate audio — and can I get isolated stems?

It generates synchronized stereo audio together with the video by default. Stem separation into five tracks (Dialogue, Score, SFX, Foley, Ambience) is a hosted/API-only feature — the weights alone don't give you separated stems.