Pika Open-Sources H3 Flex, Community Ships Turbo LoRA

Episode 1 · Pika open-sources H3 Flex, tops DesignArena open-weight video leaderboard (2026-10-30, 26 posts)

On October 30, Pika Labs released Pika H3 Flex, its first open-weight video generation model. The official Hugging Face account confirmed via repost that the model is post-trained from MiniMax H3, outperforming both H3 and H3 Max, and claimed the top spot among open-weight models on the DesignArena text-to-video (t2v) leaderboard on launch day. Everything went fully open from day one: code is hosted on GitHub under Apache-2.0, weights are on Hugging Face, and ComfyUI and diffusers are already supported.

Confirmed

• Pika H3 Flex is Pika Labs' first open-weight model, post-trained from MiniMax H3, with results surpassing H3 and H3 Max (confirmed by the official Hugging Face repost)
• Ranked #1 among open-weight models on the DesignArena text-to-video leaderboard
• Generates videos with synchronized audio in just 6 sampling steps; a 15-second clip with music can be produced in as fast as 7 seconds—faster than real time
• Supports up to 12 reference inputs (images, videos, sketches, audio)
• Iterative controls for professional creators: export of up to 5 separate audio tracks (dialogue, score, and other stems)
• Resolutions up to 4K
• Code open-sourced on GitHub under Apache-2.0, weights uploaded to Hugging Face, with support for ComfyUI and diffusers

Why it matters

• This marks Pika's symbolic shift from a closed-source product to open weights, and a textbook case of a MiniMax H3 post-train beating the original—showing post-training strategies can achieve leading results on open-weight models
• Features like audio stem export, 12 reference inputs, and 4K resolution clearly target professional creative workflows rather than casual consumer generation

• Pika Labs open-sources H3 Flex, a MiniMax H3 finetune with 6-step generation and 12 reference inputs — LatentSpacer · 2026-10-30
• Pika H3 Flex debuts as top open-weight t2v model, 15s clips in 7 seconds — multimodalart · 2026-10-30
• Pika H3 Flex confirmed as MiniMax H3 post-train, outperforming H3 and H3 Max — multimodalart · 2026-10-30
• Pika open-sources H3 Flex, a 6-step video model with synchronized audio — huggingface · 2026-10-30
• Pika H3 Flex hands-on: 15s 768p clips in 7 seconds, 80% cheaper than H3Max — HeyToha · 2026-10-30
• Pika's H3 Flex generates 15s 768p video in as little as 7 seconds — HeyToha · 2026-10-30
• Pika launches H3 Flex: 15-second videos in 7 seconds, 80% cheaper than H3Max — socialwithaayan · 2026-10-30
• Pika launches H3 Flex: 1.8x faster, up to 80% cheaper, 15s video in 7 seconds — SimplyAnnisa · 2026-10-30
• Pika launches H3 Flex: 1.8x faster than H3 Max, up to 80% cheaper, 15s clips in 7s — SimplyAnnisa · 2026-10-30
• Pika open-sources H3 Flex, claims #1 open-weight t2v model on DesignArena with 7s generation — victormustar · 2026-10-30
• Pika launches H3 Flex, top open-weight t2v model on DesignArena, 15s clip in 7s — SimplyAnnisa · 2026-10-30
• Pika's H3 Flex video model goes live: 1.8x faster, up to 80% cheaper than H3Max — FellMentKE · 2026-10-30
• Pika launches H3 Flex, a faster-than-realtime video model from $0.01/sec — ytjessie_ · 2026-10-30
• Pika releases open-source H3 Flex video model, tops DesignArena t2v board at 4K — linoy_tsaban · 2026-10-30
• Pika H3 Flex launches: faster-than-realtime video model, 1.8x quicker and 80% cheaper than H3Max — minchoi · 2026-10-30
• Pika launches H3 Flex: 1.8x faster than H3Max, 80% cheaper, up to 4K — heyshrutimishra · 2026-10-30
• Pika H3 Flex Generates 15-Second Video in 7 Seconds, Tops DesignArena T2V Board — minchoi · 2026-10-30
• Pika launches open-weight video model H3 Flex with creative controls and lower cost — cocktailpeanut · 2026-10-30
• Pika Labs drops first open-weights model, Pika H3 Flex, a MiniMax H3 post-train — multimodalart · 2026-10-30
• Pika launches H3 Flex, claims #1 open-weight t2v on DesignArena: 15s video in 7s — dr_cintas · 2026-10-30

6 more related posts →

Episode 2 · Pika H3 Flex Turbo LoRA Cuts Sampling Steps from 6 to 4 (2026-10-30, 3 posts)

A community-released fused Turbo LoRA for Pika H3 Flex reduces video sampling steps from 6 to 4 in a single plug-and-play file, with tests showing image quality nearly identical to the official 6-step output.

• Pika H3 Flex Turbo LoRA cuts video sampling from 6 steps to 4 with a single drop-in file — Illustrious_Row_9971 · 2026-10-30
• Pika H3 Flex gets a fused 4-step Turbo LoRA matching official 6-step quality — Illustrious_Row_9971 · 2026-10-30
• Pika H3 Flex Turbo LoRA: 2-Step Drafts, 4-Step Output Beats Official — _akhaliq · 2026-10-31