Pika H3 Flex — open-weight joint audio-video model post-trained on MiniMax H3

TL;DR

Pika announces Pika H3 Flex, an open-weights text-to-video+audio generation model post-trained on MiniMax H3 — Open-Weights Omni-Reference Video-Audio Generation Model Card MiniMax H3. The launch tweet claims #1 open-weight slot and #7 overall on DesignArena’s text-to-video leaderboard, outranking both its own teacher H3 and H3Max. Headline capabilities: resolutions up to 4K, exportable audio stems, and faster-than-realtime generation (a 15-second clip with soundtrack in ~7 seconds). Code is on GitHub (Pika-Labs/Pika-H3-Flex), weights on Hugging Face, with API access at $0.01/sec and 15 free generations/day for two weeks.

Key claims

Pika H3 Flex is post-trained from MiniMax H3 and open-sourced from day 1 (code + weights) [tweet thread §3-4].
Ranks #1 on DesignArena’s text-to-video board among open-weight models and #7 overall quality [tweet thread §2].
Outranks the MiniMax H3 base model and H3Max on the same leaderboard [tweet thread §2].
Supports resolutions up to 4K and exports separable audio stems — a production-iteration feature not present in the base H3 release [tweet §1].
Faster-than-realtime generation: a 15-second clip with soundtrack generated in as little as 7 seconds [tweet §1].

Method

Pika H3 Flex is a post-training derivative of MiniMax H3 — Open-Weights Omni-Reference Video-Audio Generation Model Card MiniMax H3, the 33B dense single-stream Omni-Transformer trained on packed text+video+audio tokens with MM-RoPE over (t,h,w). The tweet discloses no architectural change, no distillation recipe, and no quantitative benchmarks beyond the DesignArena rank; the “faster-than-realtime” claim is consistent with a distilled few-step derivative in the vein of FastH3 (FastVideo FastH3 V1 — 4-Step Sparse Distilled MiniMax H3 for 14x Blackwell Speedup) and TaoMate-H3 (TaoMate-H3 — Low-Latency Streaming LoRA Distillation of MiniMax H3 for Joint Audio-Video), but Pika has not confirmed which lever (step reduction, sparse attention, low-precision kernels, or capacity reduction) produces the speedup. The exportable audio stems feature is new to the H3 lineage — H3’s audio branch produces mixed stereo, and separable stem export implies additional training-time structure.

Results

DesignArena text-to-video leaderboard: #1 open-weight, #7 overall (relative to all public closed+open systems ranked there as of 2026-10-29) [tweet thread §2].
Generation time: 15-second clip with audio in ~7 seconds, i.e. ~2× realtime, on undisclosed hardware [tweet §1].
Pricing proxy: $0.01/sec on the Pika API Club, implying ~$ 0.15 for a 15-second clip at launch pricing [tweet thread §4].

Why it’s interesting

Pika H3 Flex is the first frontier video lab entry to adopt MiniMax H3 as a base and release the result openly — distinct from the systems-level accelerators already filed on the same base (FastVideo FastH3 V1 — 4-Step Sparse Distilled MiniMax H3 for 14x Blackwell Speedup, VideoDeltaNet on MiniMax H3 — Near-Lossless Hybrid Attention for Video Models, TaoMate-H3 — Low-Latency Streaming LoRA Distillation of MiniMax H3 for Joint Audio-Video, Kamimoto — Jev-gated dynamic sparsity for MiniMax H3 video generation (41.7% wall-clock reduction on RTX 4070)), which were open-source community derivatives rather than lab releases. Combined with Kling Video 2.6 — first Kling AI model with native audio (announcement) Kling 2.6 and Wan 2.6 — native multimodal video model with audio, multi-shot storytelling, starring references (Alibaba Wan announcement) Wan 2.6 shipping native joint audio in the same half-year, this closes most of the “which closed labs have joint-AV” gap flagged in Joint audio-video generation — leaving Vidu as the main holdout. The “exportable audio stems” axis is new relative to the LTX-2 / MOVA / SkyReels-V4 / H3 cluster, which all output mixed audio.