Pika H3 Flex

Pika H3 Flex is Pika Labs’ open-weights video model launched October 29, 2026: a post-train of MiniMax H3 aimed at professional iteration (faster, more controllable). At launch it ranked #1 among open-weights (about #7 overall) on DesignArena’s text-to-video board.

Blog: Introducing Pika H3 Flex · Product: pika.art/pika-h3-flex · BusinessWire

Key points

1. Post-trained H3: MiniMax H3 distilled to ~6 generation steps (vs ~50), then refined; vendor cites ~7 s DiT compute for a 15 s 768p clip (not end-to-end API latency).
2. Multimodal refs: up to 12 references (image/video/audio); text-to-video, first/last frame, reference-to-video, and more.
3. Audio stems: API can export five editable stems (Dialogue / Score / SFX / Foley / Ambience).
4. Resolution: native 480p/768p with hosted upsampling to 1080p/2K/4K.
5. Open weights: generator adapters + inference code on HF/GitHub; upsampler, stem separation, and accel stack are Pika-hosted. Creation product lands November 2.

Why it matters

Official blog + BusinessWire + high-engagement @pika_labs launch make “Pika H3 Flex” a distinct searchable product vs base MiniMax H3 / FastH3 Omni.

How to try it

1. Open the Pika H3 Flex page or call text/image/reference-to-video via Pika API Club.
2. Developers can pull open weights and inference code from Hugging Face / GitHub (local speed depends on hardware).
3. Pika’s creation product is scheduled for 2026-11-02; early promo credits/discounts follow Pika’s launch notice.

Prefer Flex for iterative, stem-aware professional video work; compare against base MiniMax H3 / H3 Max when evaluating quality vs speed.