AIREITERPika H3 Flex is the fastest useful way to iterate on short AI video with sound, but the headline "15 seconds in 7 seconds" only measures transformer compute. A real clip still takes minutes locally or under a minute through the hosted API, and anything above 768p lives behind the paywall.
The 7-second claim, corrected
The 7-second figure measures pure Diffusion Transformer compute on Pika's internal clusters, not end-to-end generation time, a distinction Pika's technical analysis makes when reporting its server numbers. The practical workflow adds text encoding, VAE decoding, audio interleaving, transport, and queueing. Expect under a minute end-to-end via the hosted API, and several minutes on a local consumer GPU.
Pika reports 3.409 seconds of server compute before serving optimizations and 2.129 seconds after, a 1.60x serving gain, for a 5-second 768p clip with audio on eight B300 GPUs (source). The kie.ai overview explicitly warns the 7-second figure for a 15-second 768p clip should not be treated as guaranteed API latency or consumer-GPU performance.
Real users confirm the gap. A compilation of peer tests found an RTX 5090 needing about 4 minutes 55 seconds for a 10-second clip at 6 steps versus 20 minutes 34 seconds stock, with picture quality nearly unchanged, while an RTX 4070 needed about 7 minutes 11 seconds for 5 seconds with notably better sound. Six steps cut sampling passes roughly 8x versus the 50-step teacher schedule, and wall-clock time roughly 3-4x on the same card in community tests: a genuine iteration win, not faster-than-realtime on your desk.
What Pika H3 Flex is
Pika H3 Flex is a post-trained MiniMax H3 distributed as two LoRA adapters, not a new video architecture. Pika fine-tuned MiniMax H3 on an internal audio-video dataset at the teacher's 50-step schedule, distilled it with projected distribution matching (PDMD) down to 6 steps after finding standard DMD unstable, then aligned it with ReFL reward learning for aesthetics and DPO for instruction following (Pika research breakdown). The Pika launch announcement positions the result as a faster, more controllable H3 derivative for professional iteration, and the official launch thread confirms open weights from day one.
What you get locally:
| Category | Pika H3 Flex local spec |
|---|---|
| Base model | MiniMax H3 (33B omni video-audio) |
| Format | 2 LoRAs: FL2VA + Ref2VA (2.77 GB FP32 each, 1.96 GB BF16 ComfyUI builds) |
| Sampling | 6 steps, lcm sampler, manual sigma schedule |
| Resolutions | 480p or 768p, 11 aspect ratios from 21:9 to 9:21 |
| Duration | 5-15 s at 24 fps (5 s = 124 frames, 15 s = 362 frames) |
| Audio | 32 kHz stereo mixed into MP4 |
| Prompt window | 1-8,000 characters |
| Launch date | October 29, 2026, open weights day one |
Pika's adapters attach to the standard MiniMax H3 base files, so existing H3 setups only add the two LoRA downloads; if H3 is new to you, budget for roughly 210 GB of base files plus 5.5 GB of adapters (footprint breakdown). The Hugging Face model repository hosts the weights, examples, and license, with Diffusers integration docs covering the two adapters and the 6-step pipeline.
FL2VA vs Ref2VA: which adapter does your shot need
FL2VA and Ref2VA map to MiniMax H3's two transformer branches, and the adapter must match both the workflow and the base checkpoint. Pairing the wrong LoRA with the wrong base is the most common setup failure in early guides like the ComfyUI setup walkthrough.
| Adapter | Inputs | Use it for | Pair with base |
|---|---|---|---|
| FL2VA | Text, optional first and/or last frame | Text-to-video, image-to-video, first-frame, last-frame, first-plus-last-frame | minimax_h3_fl2va_pruned_int8_convrot |
| Ref2VA | Text plus ordered image, video, audio references | Reference-to-video, motion transfer, character swap, reference-guided edit | minimax_h3_ref2va_pruned_int8_convrot |
Ref2VA is the distinctive one: up to 12 references per generation (per-type caps of 9 images, 3 videos, 3 audio clips, max 15 seconds combined per media type, at least one image or video), with separate control of identity from an image, motion from a video, and voice from an audio sample.
Ref2VA prompts use a structured reference format with subject_definitions, a summary, and labels like <Subject 1>, <Picture 1>, <Video 1>, <Audio 1>. ComfyUI also supports multi-shot instructions such as hard cuts, camera choreography, and dialogue in one prompt. Anyone evaluating MiniMax H3 Max API pricing for reference-heavy work should test Ref2VA first, since it is the closest open-weight route to the same shot control.
One hard rule: do not stack another speed LoRA on top. Flex is already distilled to its own 6-step latent schedule, and the ComfyUI guide warns that adding Turbo, Lightning, PTD, or 8-step Acc acceleration "breaks the output." Style or character LoRAs trained against undistilled H3 may technically load but deserve a fresh test before production use.
The split Pika kept for the API
Local output stops at 768p with a mixed stereo track. Higher resolutions, separate stems, and prompt enhancement run through Pika's hosted API.
| Capability | Local open weights | Hosted Pika API |
|---|---|---|
| Resolution | 480p, 768p | 480p, 768p, 1080p, 2K, 4K |
| Audio | Mixed 32 kHz stereo | Mixed plus 5 separate WAV stems (dialogue, score, SFX, Foley, ambience) |
| Prompt help | Manual | Prompt rewriter ($0.02 per request) |
| Reference video billing | N/A (compute only) | Input plus output seconds billed |
Hosted pricing at launch is promotional, roughly half of Pika's regular rates (pricing figures via kie.ai):
| Resolution | Promo per output second | Standard per output second | 15-s clip promo cost |
|---|---|---|---|
| 480p | $0.01 | $0.02 | $0.15 |
| 768p | $0.02 | $0.04 | $0.30 |
| 1080p | $0.035 | $0.07 | $0.525 |
| 2K | $0.04 | $0.08 | $0.60 |
| 4K | $0.08 | $0.16 | $1.20 |
Audio stems add $0.001 per audio second, and reference video inputs are billed at the generation rate for both input and output seconds. Image and audio references carry no input charge. API Club members received 15 free daily generations during the October 29-November 5 launch window (promo terms).
Pika's draft-to-final workflow (described here) generates a low-resolution draft at 480p or 768p, approves the take, then finishes the same latent seed at 1080p through 4K while paying only the price difference. A creator iterating ten drafts at 480p ($0.15 each promo) and finishing one at 4K pays the $1.50 in drafts plus the $1.05 upgrade, $2.55 total, versus $12.00 finishing all ten at 4K. For MiniMax H3 video workflows with heavy seed exploration, that structure cuts the experimentation tax.
The license catch deserves equal weight. The adapters ship under the MiniMax H3 Community License Agreement as derivatives of MiniMax H3, while only the inference code carries Apache 2.0 (license split). Prior reporting on the base license found territorial and commercial-use conditions, including excluded local deployment in the US, EU, UK, and South Korea. Pika had not confirmed additional terms at launch, so anyone planning paid client work should read the adapter repository's license file directly and treat the hosted API as the candidate compliant route in excluded regions until confirmed with Pika.
Running it locally without an H100
Pika's representative tests ran on a single 80 GB H100 with no multi-GPU sharding for a single clip (hardware notes), and the full download footprint sits around 215.5 GB. A CPU-offload flag trims VRAM at the cost of system RAM and transfer time. The rule of thumb from early guides holds: if stock H3 already runs on a given GPU, Flex runs there too, plus roughly 2 GB for the LoRA.
A detailed launch-day hardware report logged 480p 5-second runs at 66-69 GiB VRAM and 768p 15-second runs requiring CPU offload plus around 94 GiB of system memory, with a 5-second 1344x768 ComfyUI render landing near one minute. That frames the consumer reality: 480p drafts are reachable on 24 GB cards with offload, while 768p long clips are a system-RAM exercise.
"the Flex clip moves MORE. No H100. One 24GB card on a desk." — @KuGeKinbaku, comparing the same prompt and seed on an RTX 3090 Ti: Turbo 8-step at 212.61 s vs Pika H3 Flex LoRA at 6 steps in 146.81 s, about 31% faster end-to-end
The working ComfyUI recipe, per the template defaults (settings source), is narrow and worth following exactly:
| Setting | Value | Why |
|---|---|---|
| Steps | 6 | Distilled schedule; 4 steps shows popping in fine detail |
| Sampler | lcm | Matches training configuration |
| Sigmas | Manual, knots 1000/750/500/250/100/40/0 | Generic 6-step Euler is not equivalent |
| Guider / CFG | BasicGuider, CFG 1.0 | Skips the negative-prompt pass |
| LoRA strength | 1.0 | Lower values leave the model partially distilled |
| Shifts | Video 12, audio 3 | Pika reference values |
ComfyUI needs version 0.43.2 or newer for the text-to-video template (0.30.0 suffices for reference and first-last-frame templates), with three templates under Workflow > Browse Templates > Video. No custom nodes are required. For 16 GB cards, start at 480p and 5 seconds, launch with --reserve-vram 1.5, keep 64 GB of system RAM comfortable, and close other GPU-heavy apps.
Where it stands on quality
Pika H3 Flex debuted at 1293 Elo on DesignArena's text-to-video board, 7th overall and the top open-weight entry at announcement, ahead of MiniMax H3 at 1277 (leaderboard snapshot). Context from the same snapshot: Gemini Omni 1.1 Flash at 1360, Wan 3.0 at 1324, FLUX 3 Max at 1297, FLUX 3 Video at 1286, Seedance 2.5 at 1256. The 16-point lead over its teacher reads as preference optimization compensating for aggressive distillation rather than a visual revolution.
Pika's own matched benchmark at 768p favors Flex across most measured axes (Pika benchmark table):
| Metric | Pika H3 Flex, 6 steps | MiniMax H3 | Fal H3 Max |
|---|---|---|---|
| Video technical | 12.3 | 11.7 | 10.6 |
| Video aesthetic | 5.41 | 5.28 | 5.00 |
| Audio aesthetic | 7.02 | 6.85 | 6.89 |
| Audio-video alignment | 0.212 | 0.208 | 0.216 |
These are Pika's measurements, not independent verification, and Fal H3 Max edges alignment at 0.216. Pika concedes remaining weaknesses in physics, multi-scene storytelling, lighting, materials, and action-heavy reference tasks (Pika's listed limits). The working rule from early testers such as djbfilmz's 6-vs-20-step comparison: Flex for prompt and seed exploration across many clips, stock H3 at 20 steps for the final render when physics or materials carry the shot.
Prompting notes, all from the ComfyUI prompting section:
- Describe camera movement with its endpoint ("the camera slowly pushes in and stops on her face"); cut timing may drift 0.2-0.3 s.
- Stage framing through scene content ("a small figure alone in a huge empty hall") rather than bare lens terms.
- Expect fine contact interactions to stay unreliable; short on-prop text labels reportedly survive camera moves.
- The 15-second ceiling stands either way, so plan multi-shot stories as separate generations.
Should you download it or just call the API?
| Your situation | Recommended route | Reason |
|---|---|---|
| Iterating dozens of seeds or shots weekly | Download the LoRAs, draft at 480p | 6-step drafts cut exploration time ~3-4x; no per-second meter |
| Need 4K, separate stems, or prompt rewriting | Hosted API | Those features do not exist in the local weights |
| Consumer GPU (24 GB or less) | API, or 480p local drafts | 768p 15-second local runs demand offload and heavy RAM |
| Commercial delivery in US, EU, UK, or KR | API until license is cleared | Base community license may exclude local deployment there |
| Character-consistent series or motion transfer | Download Ref2VA specifically | 12-reference control is the local differentiator |
Start with a 480p draft batch locally, finish the approved seed hosted.
Pika H3 Flex FAQ
What is Pika H3 Flex?
Pika H3 Flex is Pika's post-trained MiniMax H3 released October 29, 2026 as two open-weight LoRA adapters that generate 5-15 second video with synchronized stereo audio in 6 sampling steps.
How is Pika H3 Flex different from MiniMax H3?
It keeps the MiniMax H3 architecture and base files but adds supervised fine-tuning, 6-step PDMD distillation, and ReFL plus DPO alignment, cutting sampling from 20-50 steps to 6 with slightly higher preference scores in Pika's own tests.
Can Pika H3 Flex generate 4K video locally?
No. Local output stops at 480p and 768p. Output at 1080p, 2K, and 4K plus five separate audio stems is available only through Pika's hosted API at $0.035-$0.08 per second promo rates.
What GPU do I need to run Pika H3 Flex locally?
Pika tested on an 80 GB H100 with about 215 GB of downloads. Community tests show workable 480p runs on 24 GB cards with CPU offload, while 768p 15-second renders need substantial system RAM, so most consumer setups should plan on 480p drafts.
Should I use the FL2VA or Ref2VA adapter?
FL2VA handles text-to-video, image-to-video, and first or last frame conditioning. Ref2VA handles reference-to-video tasks like motion transfer and character swaps with up to 12 combined image, video, and audio references.
Is Pika H3 Flex free for commercial use?
The weights are free to download but inherit the MiniMax H3 Community License with territorial and commercial-use conditions. Commercial teams should review the repository license file and consider the hosted API in restricted regions until Pika confirms otherwise.