What is Pika H3 Flex?
Pika H3 Flex was released on 29 October 2026 by Pika, the trading name of Mellis, Inc. It is not a new architecture: it is MiniMax H3 after three rounds of post-training (supervised fine-tuning, step distillation that cuts sampling from 50 steps to 6, and reward learning), shipped as two LoRA adapters. FL2VA handles text and first- or last-frame prompts; Ref2VA takes up to 12 reference files, at most 9 images, 3 videos and 3 audio clips.
Locally it generates 5 to 15 second clips at 480p or 768p, 24 frames a second, with 32 kHz stereo sound, across 11 aspect ratios. The hosted Pika API adds 1080p, 2K and 4K through a 1B-parameter upscaler, separate audio stems and a prompt rewriter, none of which are in the open release.
The adapters are Model Derivatives of MiniMax H3 and carry the MiniMax H3 Community License, which does not permit use of the model or its outputs in the EU, the UK, South Korea or the United States. Only the example code is Apache 2.0.
How much does Pika H3 Flex cost?
| Plan | Price | What you get |
|---|---|---|
| Text/image to video | $0.02 (480p) to $0.16 (4K) per second | $0.04 at 768p, $0.07 at 1080p |
| Reference to video | $0.06 (480p) to $0.20 (4K) per second | Input video seconds are billed too |
| Open adapters | Free | MiniMax H3 Community License; excluded in the US, EU, UK and South Korea |
Pika API per output second, text or image to video: $0.02 (480p), $0.04 (768p), $0.07 (1080p), $0.08 (2K), $0.16 (4K); reference-to-video $0.06 to $0.20. Half price at launch. Audio stems $0.001 a second. Adapters free under the MiniMax H3 Community License.
Is Pika H3 Flex worth it?
Pika H3 Flex's real offer is price and speed on Pika's API, not the open weights. Text-to-video at 768p costs $0.04 a second, half MiniMax's own $0.08 for the base model, and 1080p is $0.07. Pika claims a 15-second 768p clip in as little as 7 seconds and 1.8 times the speed of fal's H3 Max, but its footnote says that is denoising time on its servers, not end-to-end latency, and that self-hosted runs will not necessarily match it.
The quality evidence is thin. Pika cites 1293 on DesignArena's text-to-video board, seventh overall and ahead of base MiniMax H3 at 1277, a 16-point gap. Its research post shows Flex leading H3 on five automated quality metrics and trailing fal's H3 Max on the two that measure audio and text alignment; the human-preference chart has no published scores, and favours rivals on physics, lighting and multi-scene stories. Artificial Analysis had not listed Flex when we checked, where base H3 sits third at 1161.
The licence is the larger problem. Because the adapters are derivatives of MiniMax H3, they carry its terms: no use of the model or its outputs in the EU, the UK, South Korea or the United States without a separate MiniMax licence, written permission above $20 million in annual revenue, and "MiniMax H3" shown in any commercial product. For most of Pika's likely buyers the downloadable adapters are not usable, and the API is the only route.
Self-hosting is heavy where it is allowed. The base components take about 211 GB of disk, a default 480p, 5-second run used 66 to 69 GiB on an 80GB H100, and 768p at 15 seconds only fits with CPU offload and about 94 GiB of host memory. Only the H100 has been tested, and the upscaler, stems and prompt rewriter stay on Pika's servers.
Where it earns its place
A performance-marketing agency producing 1,000 ten-second product clips a month
Generate ad variants with sound from a packshot and a script
Call the Pika API at 768p for about $400 a month at the regular $0.04 rate, against about $800 on MiniMax's own H3 endpoint. Upscale the winners to 1080p by paying only the difference in rate.
A motion designer at a small branding studio
Turn a character sheet, a set image and a dance clip into one styled animation
Use the Ref2VA mode with up to 9 images, 3 videos and 3 audio clips. Reference-to-video costs $0.12 a second at 768p, and input video seconds are billed alongside output, so long motion references raise the bill.
An ML engineer at a studio in Singapore with an H100 node
Fine-tune a video model on the studio's own style
Download the adapters and the 211 GB of base components, then run one request per GPU. The licence allows this outside the excluded territories, but any derivative keeps the MiniMax terms and the 768p ceiling.
Before you commit
- Setup
- API: a Pika API key and an async job flow. Local: clone the GitHub package (pip, Conda or Docker, Python 3.12); the CLI downloads about 211 GB of pinned base components plus 5.5 GB of adapters.
- Integrations
- Pika REST API with draft-to-video upscaling, audio stems and prompt rewriter; ComfyUI templates for text, frame and reference modes; an upstream Diffusers pipeline; a Hugging Face Space.
- Data and privacy
- Adapters are under the MiniMax H3 Community License: excluded in the EU, UK, South Korea and US, written permission above $20 million revenue, "MiniMax H3" shown in commercial products. Pika's API data-retention and training terms were not published on the pages we read.
- Cost at scale
- 1,000 ten-second clips a month cost about $400 at 768p or $700 at 1080p on regular API prices; 4K is $0.16 a second. Self-hosted, a 480p 5-second run used 66 to 69 GiB on an H100 80GB.
- Limits
- 5 to 15 seconds per clip at 24 fps; 480p or 768p locally; prompts up to 8,000 characters; one request per GPU; 768p at 15 seconds needs CPU offload and about 94 GiB of host RAM.
What’s good
- $0.04 a second at 768p, half MiniMax's base H3 price
- Six sampling steps instead of 50
- Up to 12 reference files, with stereo sound
- ComfyUI and Diffusers support from launch
What’s not
- Weights barred in the US, EU, UK and South Korea by MiniMax's licence
- Quality claims are Pika's own; not yet on Artificial Analysis
- Open release stops at 768p; upscaler and stems are API-only
- Self-hosting needs an 80GB GPU and about 211 GB of disk
Best for
Teams already using MiniMax H3 who want the same model faster and cheaper through an API, and studios outside the licence's excluded regions that fine-tune on H100s.
Pika H3 Flex alternatives
Pika H3 Flex: common questions
- Is Pika H3 Flex worth it?
- Pika H3 Flex is MiniMax H3 made faster and cheaper to rent, but the open weights inherit a licence that rules out Pika's own home market. We rate it 3.8 out of 5.
- How much does Pika H3 Flex cost?
- Pika API per output second, text or image to video: $0.02 (480p), $0.04 (768p), $0.07 (1080p), $0.08 (2K), $0.16 (4K); reference-to-video $0.06 to $0.20. Half price at launch. Audio stems $0.001 a second. Adapters free under the MiniMax H3 Community License.
- What are the downsides of Pika H3 Flex?
- Weights barred in the US, EU, UK and South Korea by MiniMax's licence; quality claims are Pika's own; not yet on Artificial Analysis; open release stops at 768p; upscaler and stems are API-only; self-hosting needs an 80GB GPU and about 211 GB of disk.
- Who is Pika H3 Flex best for?
- Teams already using MiniMax H3 who want the same model faster and cheaper through an API, and studios outside the licence's excluded regions that fine-tune on H100s.
- How much does Pika H3 Flex cost for a team or in production?
- 1,000 ten-second clips a month cost about $400 at 768p or $700 at 1080p on regular API prices; 4K is $0.16 a second. Self-hosted, a 480p 5-second run used 66 to 69 GiB on an H100 80GB.
- Is Pika H3 Flex safe for company data?
- Adapters are under the MiniMax H3 Community License: excluded in the EU, UK, South Korea and US, written permission above $20 million revenue, "MiniMax H3" shown in commercial products. Pika's API data-retention and training terms were not published on the pages we read.
- What are the best Pika H3 Flex alternatives?
- The closest alternatives we have reviewed are MiniMax H3 (Hailuo 3.0), Wan 3.0, LTX-2.5 and Seedance 2.5.
Answers checked Nov 2, 2026.