Pika H3 Flex: open six-step adapters for MiniMax H3, plus audio stems

What it is. Pika H3 Flex is a post-trained version of MiniMax H3, shipped as two adapters: FL2VA (text, image, first frame, last frame or both, to video with audio) and Ref2VA (generation from references, up to 9 images, 3 videos and 3 audio clips, 12 in total). Pika announced it on Oct 29 and says it is the #1 open-weight model on DesignArena's text-to-video board. In its technical write-up, Pika also says it ranked first for human anatomy and artistic style in its own human evaluation; those are the company's numbers.

Why it's fast. Both adapters sample in six steps on a fixed schedule (knots 1000/750/500/250/100/40/0), instead of the 25 or more steps base H3 workflows usually run. The README warns that a generic six-step Euler sampler is not equivalent, so use Pika's schedule. Locally you get 480p or 768p, 11 aspect ratios, 5 to 15 seconds at 24 fps, with 32 kHz stereo audio mixed in.

What it needs. The open package is Python 3.12 on one NVIDIA GPU, tested on an H100 80GB. Pika lists about 211 GB for the base components plus about 5.5 GB for the adapters, and says a 768p, 15 s clip needs --cpu-offload on an H100. 1080p, 2K, 4K and the five editable stems (dialogue, score, SFX, foley, ambience) are only on the hosted API, which Pika's table lists from $0.01 per output second; the stems endpoint takes up to 120 s of audio or video and is billed per output stem second.

What people are reporting. ComfyUI added templates on day one, and Comfy's advice, as relayed by creators, is to keep Pika's sigma list and not stack a turbo or Lightning LoRA on top. @DeepForest_Inc reported a 10 s 768p clip on an RTX 5090 in about 4 min 55 s with Flex versus 20 min 34 s at 25 steps on base H3, with similar quality. @KuGeKinbaku measured 146.81 s for Flex versus 212.61 s for an 8-step turbo setup on an RTX 3090 Ti. @sep_is_heim found an existing 4-step LoRA faster on an RTX 4070 but preferred Flex's audio. Kijai has posted a smaller rank-56 extraction, per @ITOKtw. These are single-user reports, not benchmarks.

The licence catch. The inference code is Apache-2.0, but the adapters are MiniMax H3 derivatives under the MiniMax H3 Community License, which excludes the US, the EU, the UK and South Korea from its grant, as Ctrl+Alt+News pointed out. If you work from one of those regions, the local weights aren't licensed for you as written; Pika's hosted API is a separate commercial service. Read the licence yourself; this isn't legal advice.

Our take: six-step H3 with real audio is the most useful speed-up the open H3 ecosystem has had, and the ComfyUI templates make it easy to try on a rented GPU. For US-based shops, though, the licence makes the hosted API the practical route, and the stems endpoint (separate dialogue, score and SFX) is the part most likely to save editing time.