Developer
MiniMax

Hailuo 2.3

MiniMax's flagship video generation model, released October 28, 2025. Closed weights, text-to-video and image-to-video up to 1080p, no audio track — and still MiniMax's newest video model more than eight months later, while nine other MiniMax releases shipped in the interim.

Released
Oct 28, 2025
Type
Video Generation Model
License
Proprietary
🌐 Official Website📚 Documentation💰 Pricing & API
On this page

Overview

Hailuo 2.3 is MiniMax's flagship video generation model, released on October 28, 2025. MiniMax's announcement states that, "Building upon the Hailuo 02 model, it further enhances dynamic expression, resulting in more realistic and stable visuals," with "significant improvements in the portrayal of physical actions, stylization, and character micro-expressions." It shipped in two variants — Hailuo 2.3 and Hailuo 2.3 Fast — "across the Hailuo AI website, mobile app, and Open Platform API," alongside the rebranding of the Hailuo Video Agent into a broader Media Agent.

The commercial pitch was cost, not capability. MiniMax states Hailuo 2.3 boosts performance "while maintaining the same pricing as Hailuo 02," offering "more for the same price," while Hailuo 2.3 Fast "generates videos faster at a lower price, reducing costs for batch creation by up to 50%." The published rate card bears out the first claim exactly and the second only partly.

What makes this page unusual is what has happened since. Hailuo 2.3 is still MiniMax's newest video model more than eight months later. MiniMax's own API release notes record three video entries in total — T2V-01/I2V-01-Director on February 11, 2025, MiniMax-Hailuo-02 on June 18, 2025, and MiniMax-Hailuo-2.3 on October 28, 2025 — and then nine consecutive non-video releases: Music-2.0 and Speech-2.6 (Oct 29, 2025), M2.1 (Dec 22, 2025), Music-2.5 (Jan 16, 2026), Speech-2.8 (Jan 23, 2026), M2.5 (Feb 2026), Music2.5+ (Mar 2026), M2.7 (Mar 18, 2026), Music-2.6 (Apr 2026), and MiniMax-M3 on June 1, 2026.

Over the same window, Seedance 2.0, Kling 3.0, HappyHorse and Wan 2.7 all shipped and all now sit above Hailuo 2.3 on arena.ai's image-to-video leaderboard. MiniMax's language, speech and music lines advanced. Its video line did not.

A note on versions

There is no Hailuo 3, no Hailuo 03, and no Hailuo 2.4. MiniMax has never announced, documented, or priced any of them.

The primary sources are unambiguous. MiniMax's release notes stop at "MiniMax-Hailuo-2.3 & 2.3-Fast" on October 28, 2025. MiniMax's image-to-video API reference accepts exactly six model values — MiniMax-Hailuo-2.3, MiniMax-Hailuo-2.3-Fast, MiniMax-Hailuo-02, I2V-01-Director, I2V-01-live and I2V-01 — and no string above 2.3. MiniMax's newsroom carries no video-model announcement after it.

The confusion traces largely to hailuo3.app, an unaffiliated third-party site that markets a "Hailuo 3.0 Model," claims to be "Built on Minimax 3.0, tuned for fast iteration," names no operating company anywhere on the page, and simultaneously resells Veo 3, Sora 2, Seedance and Wan 2.5. It publishes no technical documentation for the model it names, because there is none to publish. API-aggregator blog posts advertising a "Hailuo 03" endpoint are describing the same non-existent model.

A second circulating claim deserves the same treatment. Hailuo 2.3 is widely credited with a "Physics Champion" title on WorldModelBench. That phrase appears in a press release syndicated across dozens of outlets — it does not appear in MiniMax's announcement, in MiniMax's release notes, or on WorldModelBench's own site. This page does not repeat it. Likewise, the "456-billion-parameter Mixture of Experts" figure attached to Hailuo in review articles is not published by MiniMax for any video model.

Capabilities

  • Text-to-video and image-to-video: Both modes are supported by MiniMax-Hailuo-2.3. The Fast variant is image-to-video only — it does not appear in MiniMax's text-to-video API reference.
  • 1080p output: Available at 6 seconds. Ten-second generations are capped at 768P.
  • Physical motion and micro-expression: MiniMax's stated focus. It claims the model can "render more complex character body movements with greater fluidity, naturalness, precision, and control." This is a vendor claim with no published benchmark behind it.
  • Near-photorealistic lighting: MiniMax says the model "achieves near-photorealistic visual effects in lighting direction, shadow transitions, and color tones."
  • Stylized rendering: MiniMax cites improved support for "anime, illustration, as well as special art styles like ink wash painting and game CG."
  • Media Agent: The former Hailuo Video Agent, expanded at the 2.3 launch to support "comprehensive multi-modal creation" and a "one-click video generation" flow.
  • No-charge failures: MiniMax's video-packages pricing page states that "video generation failures or videos that trigger security review will not result in a deduction."

Technical Specifications

  • API model IDs: MiniMax-Hailuo-2.3, MiniMax-Hailuo-2.3-Fast
  • Modalities: Text-to-video and image-to-video (2.3); image-to-video only (2.3-Fast)
  • Output: Silent video. No audio parameter exists in the text-to-video or image-to-video API reference.
  • Resolutions: 768P (default) and 1080P
  • Durations: 6s at 768P or 1080P; 10s at 768P only
  • Predecessor endpoints still in use: first-and-last-frame generation accepts only MiniMax-Hailuo-02; subject-reference generation accepts only S2V-01
  • Weights: Closed. No Hugging Face repository, no parameter count, no architecture disclosure
  • License: Proprietary
  • Access: MiniMax Open Platform API, the Hailuo AI web app, and the Hailuo mobile app

MiniMax publishes no parameter count, architecture description, training-data summary, or model card for Hailuo 2.3, so this page does not state one.

Use Cases

  • Short-form social video: 6-second 1080p clips at $0.49 apiece, in the framing ratios of the source image.
  • Character animation from stills: Image-to-video is the model's better-placed mode on arena.ai — rank 23 of 42, versus rank 27 on text-to-video. The two boards are calibrated separately, so their Elo numbers are not directly comparable.
  • Stylized and animated content: Anime, illustration, ink-wash and game-CG styles are MiniMax's named improvement areas over Hailuo 02.
  • High-volume batch generation: Hailuo 2.3 Fast at $0.19 per 768P 6-second clip is among the cheaper hosted video endpoints, provided you only need image-to-video.
  • Product and concept motion: Turning a product still into a short motion sequence, with sound added in post — the model produces none.
  • Storyboarding: Cheap iteration on a shot before committing to a costlier model such as Seedance 2.0 or Kling 3.0.

Performance / Benchmarks

MiniMax publishes no benchmark table for Hailuo 2.3 — no arena Elo, no WorldModelBench score, no VBench figure. Every number below is third-party.

arena.ai (formerly LMArena)

Community head-to-head voting. arena.ai's Elo scale is not comparable to Artificial Analysis's.

BoardRankEloVotes
Image-to-video23 of 421260±6229,374
Text-to-video271199±762,088

The vote count is the interesting number. On an image-to-video board carrying 1,350,288 total votes across 42 models, only xAI's grok-imagine-video-720p (462,748 votes) has been evaluated more often than Hailuo 2.3. It is the second-most-voted video model on the board — and it places 23rd.

The models ahead of it on that board include dreamina-seedance-2.0-720p (1474, rank 1), happyhorse-1.0 (1444, rank 4), wan2.7-i2v (1434, rank 5), vidu-q3-pro (1361, rank 12), kling-v3-pro (1360, rank 13) and kling-2.6-pro (1294, rank 20). The gap to the leader is 214 Elo points.

MiniMax's older models cluster just beneath: hailuo-02-pro at 1228 (rank 29), hailuo-02-standard at 1223 (rank 32), hailuo-02-fast at 1193 (rank 35). Hailuo 2.3's 32-point improvement over hailuo-02-pro is real, but it did not change the model's tier.

Artificial Analysis

Artificial Analysis's image-to-video arena publishes two boards — "With Audio" and "No Audio" — and lists no Hailuo model on either. Hailuo 2.3 is simply not evaluated there, so its absence is not evidence about audio in either direction. Dreamina Seedance 2.0 720p leads both boards, at 1189 Elo with audio and 1345 without. These are a different Elo scale from arena.ai's; do not compare the numbers.

The reading: high engagement, mid-tier output. Hailuo 2.3 is the second-most-voted model on arena.ai's image-to-video board and one of the least competitive at that board's frontier — a position that a 2025 model holding a 2026 flagship slot will tend to occupy.

Limitations

  • No audio. The API exposes no audio parameter. Every rival flagship — Sora 2, Veo 3.1, Kling 3.0, Seedance 2.0 — generates a synchronized soundtrack. Hailuo output must be scored in post.
  • 1080p is 6 seconds, full stop. Ten-second generations drop to 768P. There is no 4K path and no long-form path.
  • Eight months without a successor. MiniMax has shipped nine non-video releases since Hailuo 2.3 and zero video releases. Building a roadmap on this line means betting on a product that MiniMax's own release cadence has deprioritized.
  • Two of four video endpoints are stuck on older models. First-and-last-frame generation still accepts only MiniMax-Hailuo-02; subject-reference generation still accepts only S2V-01, a February-2025 model. Hailuo 2.3 did not inherit those capabilities.
  • Every quality claim is a vendor claim. "Near-photorealistic," "greater fluidity, naturalness, precision," "physical realism" — MiniMax published no benchmark, no evaluation protocol, and no comparison table to support any of them.
  • The "up to 50%" Fast discount does not appear in the rate card. MiniMax's own published prices show a 42.9% reduction at 768P/10s and roughly 32% at the other two output modes. And Fast cannot do text-to-video at all.
  • Closed weights, no model card. No parameters, no architecture, no training data, no self-hosting. Inference runs on MiniMax's infrastructure or not at all.
  • Consumer billing docs are stale. The Hailuo AI payment policy still describes credit entitlements in terms of "the Hailuo01 model" and "the new Hailuo02 model" — it has not been updated for 2.3.

Pricing & Access

API — pay as you go

MiniMax publishes USD per-video rates, not per-second rates.

ModelOutputPrice per video
MiniMax-Hailuo-2.3-Fast768P, 6s$0.19
MiniMax-Hailuo-2.3-Fast768P, 10s$0.32
MiniMax-Hailuo-2.3-Fast1080P, 6s$0.33
MiniMax-Hailuo-2.3768P, 6s$0.28
MiniMax-Hailuo-2.3768P, 10s$0.56
MiniMax-Hailuo-2.31080P, 6s$0.49

The legacy MiniMax-Hailuo-02 rows are identical at 768P and 1080P — which is precisely what MiniMax meant by "maintaining the same pricing as Hailuo 02." Hailuo 02 additionally offers 512P at $0.10 (6s) and $0.15 (10s); Hailuo 2.3 does not support 512P.

API — video packages

Volume buyers purchase monthly "video points" instead.

PackagePriceTermVideo pointsRPM
Standard$1,0001 month3,76020
Pro$2,5001 month9,92030
Scale$4,5001 month18,90040
Business$6,0001 month26,78050
CustomContact sales—As neededUnlimited

Point deductions: MiniMax-Hailuo-2.3-Fast costs 0.7 points (768P/6s), 1.1 (768P/10s) or 1.3 (1080P/6s). MiniMax-Hailuo-2.3 and MiniMax-Hailuo-02 cost 1 point (768P/6s) or 2 points (768P/10s and 1080P/6s).

MiniMax does not publish a USD value for a video point. Dividing package price by included points yields $0.27 per point on Standard down to $0.22 on Business — arithmetic on MiniMax's published figures, not a rate MiniMax states.

Consumer

Hailuo AI's subscription plans, from its payment policy: Standard $14.99/mo (1,000 credits), Pro $54.99/mo (4,500), Master $119.99/mo (10,000), Ultra $124.99/mo (12,000), Max $199.99/mo (20,000). A legacy Unlimited plan at $94.99/mo, purchasable only before June 18, 2025, grants unlimited credits but "can only be used with the Hailuo01 model." Membership credits expire monthly and "do not roll over."

Ecosystem & Tools

  • MiniMax Open Platform — API keys, release notes, rate cards
  • Video generation guide — the four documented modes: text-to-video, image-to-video, first-and-last-frame, subject reference
  • Image-to-video API reference — the authoritative model enum
  • Hailuo AI — the consumer web app and mobile app
  • fal — six Hailuo 2.3 endpoints across Standard (768p) and Pro (1080p) tiers
  • Replicate — hosted Hailuo 2.3, 768p and 1080p
  • Media Agent — MiniMax's one-click multi-modal creation flow, built on the Hailuo line

Community & Resources

Frequently Asked Questions

October 28, 2025. MiniMax's newsroom post is dated "2025.10.28," and its API release notes carry the matching entry: "Released MiniMax-Hailuo-2.3 & 2.3-Fast, featuring breakthroughs in body movement, facial expressions, physical realism, and prompt adherence."
No. None of these exist. MiniMax's release notes list exactly three video-model entries — T2V-01/I2V-01-Director (Feb 11, 2025), MiniMax-Hailuo-02 (Jun 18, 2025), and MiniMax-Hailuo-2.3 (Oct 28, 2025) — and the API reference accepts no model string above 2.3. The site hailuo3.app, which advertises a "Hailuo 3.0 Model" and claims to be "Built on Minimax 3.0, tuned for fast iteration," names no operating company and is not a MiniMax property; it also resells Veo 3, Sora 2, Seedance and Wan 2.5.
MiniMax publishes USD per-video rates. MiniMax-Hailuo-2.3 is $0.28 for a 768P 6-second clip, $0.56 for 768P 10 seconds, and $0.49 for 1080P 6 seconds. MiniMax-Hailuo-2.3-Fast is $0.19, $0.32 and $0.33 for the same three outputs. Failed generations are not billed: "Video generation failures or videos that trigger security review will not result in a deduction."
No. It is proprietary and closed weights. There is no Hugging Face repository, no published parameter count, no architecture description and no training-data disclosure. Access is only through MiniMax's Open Platform API or its consumer apps.
No. Neither MiniMax's text-to-video nor its image-to-video API reference exposes any audio parameter, and the pay-as-you-go rate card prices a video, not a video-plus-audio track. Third-party hosts fal and Replicate expose no audio option either.
768P at 6 or 10 seconds, and 1080P at 6 seconds only. There is no 1080P 10-second option, and 512P is available on the older MiniMax-Hailuo-02 but not on 2.3.
On arena.ai's image-to-video leaderboard it sits at rank 23 of 42 with an Elo of 1260±6 — 214 points behind ByteDance's dreamina-seedance-2.0-720p at 1474. On the text-to-video board it is rank 27 with 1199±7. It is nonetheless the second-most-voted model on the image-to-video board, with 229,374 votes.
No. MiniMax's own API release notes show nine releases after October 28, 2025 — Music-2.0 and Speech-2.6, M2.1, Music-2.5, Speech-2.8, M2.5, Music2.5+, M2.7, Music-2.6, and MiniMax-M3 on June 1, 2026 — and not one of them is a video model.
Fast is cheaper and image-to-video only: MiniMax's text-to-video API reference lists MiniMax-Hailuo-2.3 but not MiniMax-Hailuo-2.3-Fast. MiniMax says Fast reduces "costs for batch creation by up to 50%," but its own rate card shows a maximum reduction of 42.9% (768P 10s, $0.56 to $0.32) and only ~32% on the other two output modes.
MiniMax, a Shanghai-based AI company that listed on the Hong Kong Stock Exchange on January 9, 2026. Global Times reported the stock opened at HK$235.40 against an IPO price of HK$165 before "closing 109.1 percent higher," pushing the valuation past HK$103 billion (US$13.2 billion). MiniMax also builds the MiniMax-M3 language model.

Related Models

MiniMax-M3

MiniMax's native-multimodal flagship, released June 1, 2026. A 428B-parameter MoE with ~23B active parameters, a 1M-token context window built on MiniMax Sparse Attention, and open weights under a restricted community license — not the MIT terms of the M2 generation.

Kling 3.0

Kuaishou's closed-weight video generation family, launched February 5, 2026. Four SKUs (Video 3.0, Video 3.0 Omni, Image 3.0, Image 3.0 Omni), clips of 3–15 seconds with natively generated audio, and a business doing roughly USD 500 million in annualized revenue.

Seedance 2.0

ByteDance Seed's video generation model, launched in China on February 12, 2026. A unified multimodal audio-video joint generation architecture taking text, image, audio and video inputs; its global rollout was paused in March 2026 after cease-and-desist letters from Hollywood studios.

Sora 2

OpenAI's video generation model with synced audio, released September 2025. The consumer app shut down in April 2026; the API remains live until September 24, 2026.

Vidu Q3 Pro

ShengShu Technology's flagship video generation model, announced January 30, 2026. Generates up to 16 seconds of 1080p video with native audio in a single pass. Reference-to-video runs on sibling Q3 model IDs, not on Q3 Pro itself. ShengShu shipped a separate real-time model, Vidu S1, on July 3, 2026.

Wan 2.2

Alibaba's open-weights video model family, released July 28, 2025 under Apache 2.0. A 27B mixture-of-experts diffusion transformer with 14B active parameters, plus a 5B variant that runs on an RTX 4090. The later Wan 2.5, 2.6 and 2.7 releases are closed, API-only.

Explore More Models

Discover other AI models and compare their capabilities.

Browse Models← Back to Models