Activity for Coding Agents Are Becoming Supervised Business Operators

Coding Agents Are Becoming Supervised Business Operators

The strongest current signal is Codex operating real business workflows under a human approval boundary, with support automation and evidence-backed product planning close behind. The brief also filters two model announcements through their actual developer relevance and points to a timestamped shared-harness discussion and a study-worthy agent-built web project.

Kappaemme

swyx

Latent.Space

🔥 TOP SIGNAL

Codex is crossing from code generation into supervised business operations. @DevAdventur3s reports letting it cut an ad video, publish it, build the audience, and set the budget; it stopped at the Pay button and asked permission, while the operator watched the run live inside the app. swyx reports the same shape in support: a Codex computer-use agent escalated a chat and answered a blame attempt with “complete receipts,” while the humans apparently did not realize it was a bot.

Copy the boundary, not the demo: delegate reversible execution and evidence gathering, but make payment or another irreversible transition an explicit approval point. The useful control surface is the visible stop-and-ask loop.

⚡ TRY THIS

Gate the last irreversible step. Give a computer-use agent the full run-up—create the asset, publish the campaign, assemble the audience, set the budget—then require a human approval immediately before payment. Keep the live application visible rather than reviewing only the final result.
Turn feedback files into an auditable roadmap. Install npx --yes startup-feedback-engine@latest, then feed the Codex skill support tickets, interviews, surveys, reviews, sales calls, and churn notes. Inspect the resulting clusters, opportunity scores, Now/Next/Later roadmap, and evidence IDs; keep customer words, the underlying problem, what to build or validate, and what you can honestly claim as separate fields.
Queue platform/product dependencies instead of relaying them manually. swyx’s Forge dogfooding trick is to @ the relevant Codex thread and queue the next @, so a blocked product can proceed once the platform feature unblocks. It is a small but useful dependency handoff; the fully seamless platform↔product multi-agent harness remains uncommon outside genuinely multi-tenant platforms.

📡 WHAT SHIPPED

startup-feedback-engine Codex skill. A new open-source workflow claims to map recurring customer pain to evidence-backed product priorities, customer-proof candidates, objections, churn risks, and unsupported claims, with privacy-first quote and consent handling. The one-command install is npx --yes startup-feedback-engine@latest; the post says it is 100% open source, though the repository URL is only described as being in the author’s bio.
Qwen 3.8 Max entered the test queue. @agentnative_ called it a “Huge Model Release” and linked Alibaba Qwen’s announcement; Simon Willison noted Qwen 3.8 Max and MiniMax-H3 appeared within hours of each other. The evidence here is an availability signal, not a coding result—test it before drawing model-routing conclusions.
MiniMax-H3 is public on Hugging Face—but file it under media generation, not coding agents. Its model card describes a general-purpose omni-modal system that understands text, images, video, and audio and generates video with native stereo audio up to 2K and 15 seconds; the core is a 33B-parameter dense Transformer.
Cross-surface skills remain a practical product gap. Riley Brown’s complaint is that skills created in Codex do not transfer to GPT Work, even though Work spans mobile, web, and desktop. Until portability improves, keep important skill definitions in a versioned, product-independent form rather than treating a UI-local skill store as your source of truth.

🎬 GO DEEPER

Podcast clip — Codex and ChatGPT Work, 00:07:17–00:12:07; focus on 00:09:49–00:10:23. The episode’s shared-harness chapter is the useful context for the portability problem: Akshay Nathan says Codex and Work use the same underlying harness, with improvements for plugins, computer use, and artifacts, while the UX differs—Codex exposes diffs and file edits that Work hides.
Repo to study — thebuggeddev/anatomy. This is a concrete staged asset pipeline: generate a design image, create organ images, convert them to 3D models with Tripo, then give Codex a master prompt plus the design and models. Iteration—not a one-shot prompt—cut individual models from roughly 120–150 MB to 2–5.5 MB and the total asset load from about 900 MB to 28.6 MB, with on-demand loading; the post links both a live demo and the code.

Editorial take: The alpha is shifting from “make the agent run longer” to “give it enough surface area to finish, then make the stop condition and evidence trail explicit.”

Coding Agents Are Becoming Supervised Business Operators

Research extraction

Direct answer: MiniMax-H3 is an omni-modal video/audio generation system, not a text LLM, with a 33B-param core. The card covers model size, architecture, licensing, and local deployment, but contains no stated context length, pricing, coding benchmarks, or tool-use capabilities — those are explicit gaps.

1. Model size and precision

H3-Omni-Transformer is a 33B-parameter dense, single-stream Transformer; ~13B parameters are in AdaLN branches whose outputs can be precomputed/cached, so they are not needed for inference-only deployment; full weights are released for fine-tuning.
Card metadata lists "33B params" and tensor types "F32 BF16".

2. Architecture and capabilities

System = H3-Context-IR (hosted preprocessing/orchestration, critical for output quality, not in this open-source release; API provided) → H3-Base (768p video+audio generation) → H3-Regenerate-2K (2K in-context regeneration, not yet open-sourced; API provided).
Output: 4–15 s video at 24 FPS, default 768p (2K via regeneration), 32 kHz stereo audio, stable dialogue support in 11 languages.
H3-Base: text via H3-Encoder, visuals via H3-Encoder + H3-VisualVAE, audio via H3-AudioVAE, packed into one sequence with RoPE; H3-Omni-Transformer jointly predicts video and audio latents.
H3-Encoder uses full pretrained Qwen3-VL-32B weights (hidden states from layer 50); the H3 tokenizer config with special tokens such as is required.
H3-VisualVAE: spatial 16×, temporal 4×, 24 channels (f16t4d24); tokens patchified 1×2×2 → effective spatial 32×.
H3-AudioVAE compresses 32 kHz audio to 40 Hz latent tokens per channel (shared encoder/decoder across stereo channels).
Attention and FFN have no modality-specific structure; 3D MM-RoPE encodes (t,h,w).
Sparse attention is natively supported but absent from the initial release (inference is full-attention only).

3. Reasoning / tool-use

No tool-use or agentic capabilities are described. The only reasoning reference is internal to H3-Context-IR, whose workflow includes "complex logical reasoning" for interpreting multimodal instructions.

4. Context length — gap

Not specified in the card. Documented limits are output duration (4–15 s), reference-input caps (≤9 images, ≤3 video clips 2–15 s, ≤3 audio clips, max 12 files total), and API token-usage examples (e.g., 39,299 total tokens for a 5 s Ref2VA task); none define the H3 context window.

5. Licensing

Released under the MiniMax H3 Community License Agreement (Q&A doc linked); guardrails do not affect licensee obligations on lawful use and use restrictions.

6. Local inference / setup

Two BF16 CFG-distilled task checkpoints: H3-Base FL2VA (t2va/fl2va) and H3-Base Ref2VA (ref2va), each with processor, tokenizer, text encoder, transformer, VisualVAE, AudioVAE.
Repo hosts original checkpoints and diffusers format side by side: hf download MiniMaxAI/MiniMax-H3 --include "FL2VA/*" "Ref2VA/*" for SGLang/vLLM; diffusers can use ModularPipeline.from_pretrained("MiniMaxAI/MiniMax-H3").
Recommended frameworks: SGLang, vLLM, diffusers, ComfyUI (T2V/R2V templates).
SGLang reference: sglang serve --model-path MiniMaxAI/MiniMax-H3 --num-gpus 4 --ulysses-degree 4 --performance-mode speed --model-variant fl2va|ref2va.
Full 2K reproduction combines local SGLang with hosted H3-Context-IR and H3-Regenerate-2K APIs.

7. Coding benchmarks / pricing — gaps

No coding benchmarks or prices appear anywhere in the card; it only links API docs (platform.minimax.io / platform.minimaxi.com) and apps (hailuoai.video, hub.minimax.io).

Hugging Face

Research extraction

Latent.Space

latent

Episode: Codex from 0 to 10M Users: Building ChatGPT Work — Akshay Nathan, OpenAI

Transcript/chapter timestamps: Full transcript opens at L140; chapter timestamps (L100-L136) cover: intro -> 00:01:33 joining OpenAI -> 00:02:40 enterprise AI adoption -> 00:05:28 why ChatGPT Work -> 00:07:17 Codex vs ChatGPT Work and shared agent harness -> 00:12:07 why OpenAI merged agent experiences -> 00:16:24 models/reasoning defaults -> 00:20:26 artifacts/agentic spreadsheets -> 00:24:22 Sites replacing decks/spreadsheets -> 00:30:08 designing an agent that can build almost anything -> 00:34:28 developer agents to knowledge work -> 00:36:07 power-user advice/reviews -> 00:40:41 internal AI memes and 10M-user launch -> 00:44:39 OpenClaw, personal agents, ChatGPT as OS -> 00:50:24 sub-agents/Ultra -> 00:54:39 memory/Chronicle -> 01:00:19 AI reshaping product dev/roles -> 01:03:15 ideas/taste/LLM idea generation -> 01:04:42 productivity, quality at-bats, motion vs progress.

Most useful short segment (00:09:49-00:10:23, L196-L202): Akshay describes the shared harness behind Codex and ChatGPT Work: 'the harness is the same... we made improvements to the harness to make it good for knowledge work, especially as it relates to plug-ins or computer use or artifacts... but the underlying harness and capabilities should be the same.' He contrasts the UX: in Codex mode you see Git diffs and file edits (e.g., a retirement-calculator spreadsheet), while in Work that Git detail is hidden; sandbox defaults are also opinionated per mode.

Routing signals:

ChatGPT Work grew out of Codex unexpectedly taking off among non-developers inside OpenAI - staff in strategic finance/marketing used it and felt they had a 'superpower'; this drove the Merge/Super App decision.
Merge rationale: AI blurs boundaries between coding, strategy, marketing, and other knowledge work, so OpenAI kept primitives (e.g., plugins) unified rather than boxing users into separate experiences.
Scale: OpenAI reported 10M ChatGPT Work + Codex users combined within two weeks of the July 9 launch, with Codex MAU up >10x since Jan 2026; separately, Codex had 5M+ weekly active users, knowledge workers were ~20% of the base in June 2026 and growing >3x as fast as developers.
'Computer use' is positioned as a capability OpenAI is bringing to knowledge workers, alongside artifacts and better context gathering.

Codex from 0 to 10M Users: Building ChatGPT Work — Akshay Nathan, OpenAI

swyx

Firsthand: @swyx ran a Codex computer-use (CUA) agent on a real support chat — the agent escalated the issue for faster resolution, and when the support agent blamed his side, the bot "replied with complete receipts"; the humans "have no idea they are talking to a bot" (screenshots in thread). It's part of a running list of Codex CUA wow moments he's collecting ahead of his computer-use podcast .
Scale/status: Codex and ChatGPT Work crossed 10M paid users; paid users get a fresh daily usage reset, per @thsottiaux's announcement .
Pod signal: @swyx recorded a Latent Space episode with @akshaynathan_, who leads Productivity engineering, covering Codex + ChatGPT Work + the 10M milestone; episode is live at https://www.latent.space/p/chatgpt-work. His opinion (not fact): "Work + GPT 5.6 is the most company defining launch since og chatgpt itself," and with @AriX's computer use it "is gonna reach >1B users worldwide" .

in prep for our computer use pod, gonna store a running list of codex cua wow moments. here it is dealing with support chat for me to esc… 10M! New day, new usage reset for paid users of Codex and ChatGPT Work. Lands in the next hour. Enjoy. ![](https://pbs.twimg.com/media/HN… recorded Codex + ChatGPT Work + 10M user milestone pod with [@akshaynathan_](https://x.com/akshaynathan_) who leads Productivity engineer… [@akshaynathan_](https://x.com/akshaynathan_) [@AriX](https://x.com/AriX) live now! [https://www.latent.space/p/chatgpt-work](https://www…

Simon Willison

MiniMax announced MiniMax-H3 is now publicly available on Hugging Face . Simon Willison flagged that the release came within hours of Qwen 3.8 Max .

MiniMax-H3 Is Now Publicly Available [https://huggingface.co/MiniMaxAI/MiniMax-H3](https://huggingface.co/MiniMaxAI/MiniMax-H3) [![Video]… Qwen 3.8 Max and MiniMax-H3 within hours of each other [https://x.com/minimax_ai/status/2084106804032872591](https://x.com/minimax_ai/sta…

swyx

@swyx is developing Forge and dogfooding it by hosting all his projects on it, which makes him constantly bounce between platform and product work . He shared an OpenAI Codex trick: you can @ a thread and queue up the @, so if a project is blocked on a platform feature, you can premove the project to proceed once the platform is unblocked . He observes he wasn't strictly necessary in the process, and argues an even better multiagent harness would seamlessly orchestrate work back and forth between platform and project — though that's uncommon unless you're building a real platform with multiple product/application tenants . He previously started work on Forge agents .

one way i'm developing Forge ([https://x.com/swyx/status/2083654369095156219](https://x.com/swyx/status/2083654369095156219)) is to use i… started work on forge agents today ![](https://pbs.twimg.com/media/HOqgmyLa0AApDDD.jpg) [https://x.com/swyx/status/2080750437133901925](h…

Simon Willison

Alibaba announced Qwen3.8 is launching and going open-weight . The model has 2.4T parameters and is claimed to be compatible with leading frontier AI models, second only to Fable 5 . Qwen3.8-Max-Preview is already available on Alibaba's Token Plan, Qoder, and QoderWork . Simon Willison notes the current model is actually qwen3.8-max, while the earlier announcement called it Qwen3.8-Max-Preview . The Qwen Foundation Model Team opened a new @QwenDevs account and is hosting an AMA on X .

Qwen3.8 is launching and going open-weight soon!🌐 With a massive 2.4T parameters, this model is continuously evolving. We believe it’s on… [@Alibaba_Qwen](https://x.com/Alibaba_Qwen) Looks like this is qwen3.8-max - that older model was qwen3.8-max-preview git init qwen\_devs README.md: Hey 👋 We're the folks from Qwen Foundation Model Team. Since we finally have an account...an AMA? ![](http…

Andrej Karpathy

@karpathy ran Opus 5 on the first paragraph of Lord of the Rings with a 1M-token budget (~$10) and a three.js render request; the model autonomously worked ~2 hours and wrote 5,500 lines of procedural code. He calls the result janky, but notes it's a task no human would bother doing, and points toward on-demand 'ephemeral GTA of X' worlds .

He flags a key coding-agent weakness: LLMs can't natively perceive videos or play games, so Opus 5 had to slowly take screenshots to audit its work, producing jank — an example of the self-verification gap for long-running agentic tasks .

Karpathy uploaded the source so it's playable/forkable at karpathy.ai/lotr-movie/, and links Simon Willison's 'pelican on a bicycle' piece (simonwillison.net/2025/Jun/6/six-months-in-llms/) as the inspiration .

We're starting to leave the territory where you'd test an LLM by e.g. "create an svg of pelican on a bicycle". As one idea to generalize … More on the pelican on the bicycle test from [@simonw](https://x.com/simonw): [https://simonwillison.net/2025/Jun/6/six-months-in-llms/](…

Greg Brockman

@thebuggeddev (firsthand) vibe-coded a 3D human anatomy app: created a design image with GPT Image 2.0, generated each organ image, converted images to 3D models with Tripo AI, then wrote a master prompt for Codex and fed it the prompt, design image, and all 3D models; Codex built the Three.js app using GPT 5.6 Sol . Initial models were 120–150 MB each (16fps); after iterations Codex compressed each to 2–5.5 MB, cutting total assets from ~900 MB to 28.6 MB with on-demand loading while preserving quality, and it also generated anatomical illustrations and interactive hotspot markers . The build was explicitly step-by-step, not one-shot . Live demo and repo are public . OpenAI president Greg Brockman shared it as 'chatgpt for building interactive educational tools' .

Only if education could be this interactive ❤️‍🔥 I've had a looong wish to build something genuinely useful through vibe coding, and I final… chatgpt for building interactive educational tools [https://x.com/thebuggeddev/status/2083884856531177942](https://x.com/thebuggeddev/sta…

swyx

Secondhand aphorism from Patrick McKenzie ('Patio11'), shared by @swyx: "Patio11's Law of Agents" — you are insufficiently ambitious about what agents can do to improve every part of your workflow, and this holds even when you explicitly try to account for the law itself . Practical takeaway for agent practitioners: audit every step of your dev workflow for agentification, and expect your first-pass estimates of agent upside to be too low.

Patio11's Law of Agents: you are insufficiently ambitious about what agents can do to improve -every- part of your workflow, even when ta…

Jediah Katz

@Ananth7e claims that gpt-5.6 luna on max reasoning gives basically the same intelligence as sol on medium, at 25x cheaper than sol ("luna max = sol medium") . @jediahkatz challenges the comparison, asking to "show the latency" .

luna max = sol medium gpt-5.6 luna on max reasoning gives you basically the same intelligence as sol on medium, at 25x cheaper than sol. … Now show the latency [https://x.com/ananth7e/status/2083412521504194718](https://x.com/ananth7e/status/2083412521504194718)

Riley Brown

OpenAI's Codex and GPT Work do not share skills, per @rileybrown: a skill built in Codex cannot be used in GPT Work, and he wants skills to be global across both tools . GPT Work runs on mobile, web, and desktop apps .

My only BIG complaint with GPT Work and Codex is that the skills don't transfer. When I create a skill on Codex I want to be able to use …

Riley Brown

Alibaba Qwen released Qwen 3.8 Max, called a "Huge Model Release" by @agentnative_ (Riley Brown's agent-focused podcast); Riley Brown says he will test the model tonight . The post links to Alibaba Qwen's official announcement tweet . No capabilities, benchmarks, or pricing details are provided yet.

New model. Will be testing tonight. [https://x.com/agentnative_/status/2084105883026342088](https://x.com/agentnative_/status/20841058830… Huge Model Release: Qwen 3.8 Max. Let the testing begin. ![](https://pbs.twimg.com/media/HOw7QVCWkAArdhD.jpg) [https://x.com/alibaba_qwen…

Greg Brockman

Greg Brockman (@gdb, President & Co-Founder @OpenAI) shared an open-source Codex skill by @Kappaemme1926 that turns customer feedback — support tickets, interviews, surveys, reviews, sales calls, and churn notes — into evidence-backed product priorities and credible customer proof . The skill clusters feedback with source traceability, scores product opportunities, produces a Now/Next/Later roadmap, surfaces customer-proof/case-study candidates, objections, churn risks, and unsupported claims, and keeps customer quotes, underlying problems, build/validate candidates, and honest claims separate . Install: npx --yes startup-feedback-engine@latest; 100% open source, repo in the author's bio, with a demo video in the post . Shared by @gdb as "codex for customer feedback -> roadmap" .

CODEX SKILL THAT TURNS CUSTOMER FEEDBACK INTO A ROADMAP! Most feedback analysis stops at positive or negative. I made a Codex skill that … codex for customer feedback -> roadmap [https://x.com/Kappaemme1926/status/2083957246674042890](https://x.com/Kappaemme1926/status/208…

Greg Brockman

In a first-hand account, @DevAdventur3s let OpenAI's Codex app run ads for their business: it cut the video, published it, built the audience, set the budget, then stopped at the Pay button and asked permission before completing the financial transaction . They watched the whole process live inside the app . OpenAI president Greg Brockman shared the post as an example of "Codex for helping operate your business" . The account demonstrates a human-in-the-loop pattern: the agent handles creative and execution work but halts before irreversible payment.

I let OpenAI's Codex app run ads for my business. It cut the video, published it, built the audience, set the budget then stopped at the … Codex for helping operate your business [https://x.com/devadventur3s/status/2084054616946176324](https://x.com/devadventur3s/status/20840…

Hugging Face

Open on huggingface.co Open 19 hours ago

1 of 27 highlights

MiniMax

(opens in new tab) (opens in new tab) (opens in new tab)
(opens in new tab) (opens in new tab) (opens in new tab) (opens in new tab) (opens in new tab)

MiniMax H3

System Overview

MiniMax H3 is a general-purpose, omni-modal generative system. It supports unified understanding of multimodal contexts composed of text, images, video, and audio, and can generate video with native stereo audio at resolutions up to 2K and durations of up to 15 seconds. Thanks to its task-generalization-oriented system design, H3 already possesses broad multimodal context understanding and generation capabilities at the pre-training stage, enabling outstanding performance in following complex multimodal instructions.

H3 supports the following input and output specifications:

Category	Specification
Output duration	4–15 seconds
Output aspect ratio	Supports a wide range of aspect ratios, including but not limited to 21:9, 16:9, 4:3, 1:1, 3:4, and 9:16
Output resolution	Supports various resolution dimensions. The shorter side is set to 768 pixels by default. 2K \| generation can be achieved with H3-Regenerate-2K
Output frame rate	24 FPS
Output audio	32 kHz stereo
Supported dialogue languages	Stable support for 11 languages: Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, and Spanish. Additional languages are also supported to varying degrees

Model Variants and Input Specifications

Model Variant	Input Mode	Specifications
H3-Base-FL2VA	First-and-last-frame mode	Supports zero, one, or two input images. - No image input: Text-to-video mode - One image input: First-frame-to-video or last-frame-to-video generation - Two image inputs: First-and-last-frame-to-video generation
H3-Base-Ref2VA	Omni-reference mode	Supports multi-modal reference inputs: - Images: ≤ 9 images - Videos: ≤ 3 clips; each clip must be 2–15 seconds long; total duration ≤ 15 seconds - Audio: ≤ 3 clips; audio must be accompanied by image or video input and cannot be used as the sole input; each clip must be 2–15 seconds long; total duration ≤ 15 seconds - Mixed inputs: Maximum number of files across all input types is 12

The complete H3 system consists of the following three modules:

H3-Context-IR: As inputs become increasingly complex, we build a dedicated system to deeply understand and refine the input multimodal instructions, then convert them into a form that H3 can readily understand—the Context Intermediate Representation—for generation. H3-Context-IR is critical to the quality of the final output, so we strongly recommend incorporating it into your generation pipeline or following the “Prompting Guidance” to build your own context-processing system.
H3-Base: Generates audio and video based on the H3-Context-IR output, producing results at 768p resolution.
H3-Regenerate-2K: Feeds the 768p result together with the original context back into H3 to regenerate the output at 2K resolution. This process leverages both H3’s powerful generative capabilities and the rich information contained in the original context, enabling it to produce high-resolution outputs with more accurate details and greater visual fidelity.

Online API

Use MiniMax-H3 directly via API.

Global: platform.minimax.io (opens in new tab) | CN: platform.minimaxi.com (opens in new tab)

Online App

Use MiniMax-H3 directly via App.

WebApp Global: hailuoai.video (opens in new tab) | CN: hailuoai.com (opens in new tab)
Desktop Global: hub.minimax.io (opens in new tab) | CN: hub.minimaxi.com (opens in new tab)

Model Architecture

H3-Context-IR

H3-Context-IR is a hosted preprocessing and orchestration system designed for free-form multimodal inputs.

It interprets the relationships among text, images, audio, and reference videos, as well as how these materials relate to the intended generation output. Its internal workflow includes instruction parsing, cross-modal association, temporal understanding, and complex logical reasoning.

H3-Context-IR serializes its understanding of the context into a structured representation accepted by H3-Base. Without deviating from the user’s original intent, it may also supplement missing or underspecified semantic details where appropriate.

Because H3-Context-IR relies on a multi-stage workflow and multiple hosted models and services, it is not included in this open-source release. We provide an API that enables users to reproduce the behavior of the official workflow. We also provide detailed tutorials, and developers can follow the Prompting Guidance to build their own preprocessing systems.

For detailed usage instructions, see Recommended Workflow — Full 2K Workflow.

Safety Guardrails

User-submitted text, images and videos, as well as enhanced prompts, are subject to automated moderation. Content suspected of being unlawful, pornographic, or infringing third-party rights may be blocked. We use industry-standard filtering measures but cannot eliminate false positives or false negatives. These guardrails do not affect the Licensee’s obligations under the MiniMax H3 Community License, especially those relating to lawful use and use restrictions.

H3-Base#### Architecture Overview

H3-Base encodes different modalities using their corresponding encoders or VAEs and organizes the encoded representations into a unified packed multimodal sequence. RoPE is used to capture the necessary spatial and temporal relationships among tokens before the entire sequence is passed to the H3-Omni-Transformer.
Specifically, text is encoded by the H3-Encoder; visual inputs are encoded by both the H3-Encoder and the H3-VisualVAE; and audio is encoded solely by the H3-AudioVAE.
The H3-Omni-Transformer jointly predicts video and audio latents, which are then decoded into video and stereo audio, respectively.
To reduce the computational cost of long multimodal sequences, H3 natively supports sparse-attention training and inference. The initial open-source release provides inference with full attention only. Our sparse-attention implementation will be released in a future update.

H3-Encoder

The H3-Encoder uses the full pretrained weights of Qwen3-VL-32B and provides the hidden states from its 50th layer to the H3-Omni-Transformer.
We add several special tokens, such as <d>, to the tokenizer configuration. When using H3, the tokenizer and associated configuration files provided in the H3 repository are required.

H3-VAE

H3 uses separate visual and audio latents to represent their respective modalities.

H3-VisualVAE

H3-VisualVAE is a temporally causal video autoencoder with a spatial compression factor of 16×, a temporal compression factor of 4×, and 24 latent channels, denoted as f16t4d24. We apply several latent-space optimization techniques to jointly improve reconstruction quality and latent learnability.
Before being passed to the H3-Omni-Transformer, the visual latents are further patchified with a patch size of 1 × 2 × 2 along the (time, height, width) dimensions. As a result, the visual tokens entering the Transformer have an effective spatial downsampling factor of 32×, while the temporal downsampling factor remains 4×.
The latent space of H3-VisualVAE is optimized for both reconstruction quality and ease of learning by the generative model. After training its encoder, we additionally train a ViT-based decoder to reduce decoding costs and further improve reconstruction quality.

H3-AudioVAE

H3-AudioVAE uses the same encoder and decoder for both the left and right audio channels while processing each channel independently. The decoded channels are then recombined, enabling stereo audio input and output.
For each channel, H3-AudioVAE compresses 32 kHz audio into a sequence of latent tokens with a temporal rate of 40 Hz.
Inspired by VA-VAE, we optimize the latent space to preserve audio reconstruction quality while making it easier for the generative model to learn.

H3-Omni-Transformer

For scalability and generalization, we adopt a relatively simple Transformer block design. H3-Omni-Transformer is a 33B-parameter dense, single-stream Transformer, with approximately 13B parameters residing in AdaLN-related branches. Because the AdaLN modulation outputs can be precomputed and cached, these parameters do not need to be loaded for inference-only deployment. We release the complete model weights to support further development, including fine-tuning.
Neither the attention layers nor the FFN layers contain modality-specific structures. Modality-specific parameters are confined to the input/output layers and the AdaLN branches. In particular, modality-specific AdaLN improves generation quality with relatively low additional training and inference costs.
The model uses three-dimensional Multimodal Rotary Position Embeddings (MM-RoPE) to represent positional relationships across the temporal and two spatial dimensions, (t, h, w).
During the final stage of training, we introduce native sparse attention to reduce the computational cost of long sequences. The sparse-attention implementation is not included in the initial open-source release and will be published separately in a future update.

H3-Regenerate-2K

For H3’s 2K-resolution output, instead of using a conventional dedicated super-resolution module, we use the H3 base model to regenerate its own low-resolution result through an in-context manner.
This approach provides two advantages: (1) the regeneration process can reuse the generative capabilities of H3 base model to the greatest extent possible; and (2) the in-context format can reuse the original multimodal context when producing high-resolution output, allowing it to recover information that conventional super-resolution methods would otherwise have to “guess,” such as small text and fine details.
In-context regeneration is also an example of task generalization.
Due to the complexity of the system, this module is not yet open-sourced. We will release it once it is ready. We provide an API for validating the official results; see “Full 2K Workflow” below.

Recommended Workflow

To help the community deploy MiniMax H3 correctly, we provide two validation methods.

Since the complete H3 system consists of three modules—H3-Context-IR, H3-Base, and H3-Regenerate-2K—the “Full 2K Workflow” provides an end-to-end validation pipeline for 2K output, combining the Open Platform API with a locally deployed H3-Base. The “Local Deployment of H3-Base” section provides a method for validating 768p output using only a locally deployed H3-Base.

In addition, the “Prompting Guidance” section provides a detailed tutorial to help the community develop their own prompting systems.

Local Deployment of H3-Base

MiniMax H3 is released as two task-specific checkpoints. Each checkpoint contains a specialized Omni Transformer Model together with the required processor, tokenizer, text encoder, Visual VAE, and standalone Audio VAE components.

Checkpoint	Supported Tasks	Input Conditions	Output	Precision
MiniMax-H3 Base FL2VA	Text-to-Audio-Video (`t2va`), First/Last-Frame-to-Audio-Video (`fl2va`)	Text; optional first frame, last frame, or both	Video and audio	BF16
MiniMax-H3 Base Ref2VA	Reference-to-Audio-Video (`ref2va`)	Text with reference images, videos, and/or audio	Video and audio	BF16

The released checkpoints are CFG-distilled Omni Transformer model weights.

Each checkpoint is distributed as a self-contained Hugging Face-style repository with the following components:

<TASK>/
├── model_index.json
├── processor/
├── tokenizer/
├── text_encoder/
├── transformer/
├── visual_vae/
└── audio_vae/

Download the model. The repository hosts the original checkpoint (FL2VA/, Ref2VA/) and the diffusers format side by side, so scope the download to what your framework needs:

# Original checkpoint, both task families (SGLang, vLLM):
hf download MiniMaxAI/MiniMax-H3 --include "FL2VA/*" "Ref2VA/*" --local-dir MiniMax-H3

# Or a single task family:
hf download MiniMaxAI/MiniMax-H3 --include "FL2VA/*" --local-dir MiniMax-H3

diffusers users do not need a manual download: ModularPipeline.from_pretrained("MiniMaxAI/MiniMax-H3") fetches exactly the components it needs. See the diffusers documentation (opens in new tab) for loading recipes.

We recommend the following inference frameworks to serve the model:

Sglang Deployment

Here we use sglang as a deployment example. See the MiniMax-H3 deployment guide (opens in new tab) for additional deployment configurations.

FL2VA:

sglang serve \
  --model-path MiniMaxAI/MiniMax-H3 \
  --num-gpus 4 \
  --ulysses-degree 4 \
  --performance-mode speed \
  --host 0.0.0.0 \
  --port 30010 \
  --model-variant fl2va

Ref2VA:

sglang serve \
  --model-path MiniMaxAI/MiniMax-H3 \
  --num-gpus 4 \
  --ulysses-degree 4 \
  --performance-mode speed \
  --host 0.0.0.0 \
  --port 30011 \
  --model-variant ref2va

Reproducible 768p cases

The following three use cases T2VA, FL2VA, and Ref2VA demonstrate how to reproduce MiniMax-H3 video-audio generation.

Use case	Request	Result
T2VA	View script (opens in new tab)	t2va.mp4 (opens in new tab)
FL2VA	View script (opens in new tab)	fl2va.mp4 (opens in new tab)
Ref2VA	View script (opens in new tab)	ref2va.mp4 (opens in new tab)

Full 2K-Workflow

This section explains how to combine a locally deployed SGLang service with the official H3-Context-IR and H3-Regenerate-2K APIs to reproduce the quality of 2K videos generated directly by the MiniMax API. Before you begin, configure the SGLang endpoint and your MiniMax API credentials:

# URL of your SGLang deployment
SGLANG_DEPLOYMENT_URL="<sglang-deployment-url>"

# MiniMax API endpoint (choose one)
# CN
MINIMAX_API_BASE="https://api.minimaxi.com"
# Global
# MINIMAX_API_BASE="https://api.minimax.io"

# API token obtained from the MiniMax platform
TOKEN="<token>"

MiniMax platform:

API docs:

Create H3-2K: use /video-generation-v2-create EN-docs (opens in new tab), CN-docs (opens in new tab)
H3-Context-IR：use /video-generation-v2-h3-context-ir EN-docs (opens in new tab), CN-docs (opens in new tab)
H3-Regenerate-2K：use /video-generation-v2-regeneration EN-docs (opens in new tab), CN-docs (opens in new tab)

The examples below encode local H3-Base output files as Base64 Data URLs. For production use, uploading the video to a publicly accessible URL and passing that URL as base_video is recommended.

For each case below, we provide reference outputs at both 2K and 768p generated directly through the Open Platform API, making it easier to validate the results.

case-T2VA

Type: Text-to-video
Duration: 10 seconds
Aspect ratio: 16:9

stage	request	result
H3-Context-IR	View script (opens in new tab)	json { "task": { "id": "<task_id>", "model": "MiniMax-H3", "status": "succeeded", "created_at": "<created_at>", "updated_at": "<updated_at>", "content": { "prompt": "integrated_multimodal_description: [Shot 1] Cinematic, medium wide shot, pushing in slowly. In the cavernous, dimly lit bridge of a starship, sleek metallic consoles with glowing amber displays flank a massive, curved observation window. A female captain, in her late 40s with an athletic build and short silver-streaked black hair, stands in the center midground. She wears a structured, high-collared dark navy military tunic with silver chest insignias. Her back is to the camera, silhouetted against the cool, ambient starlight pouring through the thick glass. She stands perfectly still with her hands clasped tightly behind her back. Outside the window, a massive armada of jagged, dark grey dreadnoughts hovers in tight formation against a deep purple space nebula. The fleet's massive rear thrusters begin to glow with an intense, escalating bright blue light. [Shot 2] At 00:04.500, the camera cuts to a close-up of the captain's face and shakes strongly. The brilliant blue-white light from the fleet's gathering energy reflects vividly in her dark eyes. Suddenly, a blinding white flash floods through the window, completely washing out the background as the fleet jumps to hyperspace. The sheer spatial force violently jolts the bridge, causing the captain from Shot 1 to stagger slightly forward, her shoulders tensing as she visibly braces herself against the physical tremors. As the intense white light fades abruptly, leaving only the dim, empty expanse of the purple nebula reflected on her starkly lit skin, her jaw clenches, and she slowly closes her eyes in the newly emptied space.\noverall_soundscape: A low, resonant hum of the ship's ambient life support systems serves as the baseline, soon drowned out by an audible, escalating, high-pitched electronic whine as the fleet outside charges its hyperdrives. A massive, deafening, bass-heavy boom and sharp crackle erupts during the blinding flash, accompanied by the loud metallic creaking, rattling, and deep thuds of the bridge's bulkheads vibrating under immense physical stress. The intense roaring impact then cuts abruptly back to a hollow, echoing room tone, leaving only the faint, steady hum of the isolated bridge.\nnon_diegetic_music: Cinematic space-opera orchestral score, slow tempo, featuring a solitary, mournful French horn melody over deep, sustained string dissonances that build rapidly in volume and intensity, swelling to a massive orchestral peak before snapping immediately into silence right after the jump." }, "duration": 10, "usage": { "total_tokens": 8565, "prompt_tokens": 5650, "completion_tokens": 2915 }, "ratio": "16:9", "task_type": "h3_context_ir", "modality": "text" } }
H3-Base	View script (opens in new tab)	t2va.mp4 (opens in new tab)
H3-Regenerate-2K	View script (opens in new tab)	t2va_2k.mp4 (opens in new tab)
Reference 2K result by directly calling Open Platform API	View script (opens in new tab)	h3_direct_2k.mp4 (opens in new tab)
Reference 768P result by directly calling Open Platform API	View script (opens in new tab)	h3_direct_768p.mp4 (opens in new tab)

case-I2VA

Type: First-frame image-to-video
Duration: 8 seconds
Aspect ratio: adaptive

stage	request	result
H3-Context-IR	View script (opens in new tab)	json { "task": { "id": "<task_id>", "model": "MiniMax-H3", "status": "succeeded", "created_at": "<created_at>", "updated_at": "<updated_at>", "content": { "prompt": "For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced.\n\nintegrated_multimodal_description: [Shot 1] This is a live-action, cinematic shot with a shallow depth of field. The camera holds a perfectly static shot throughout the entire eight-second duration, capturing a cozy family gathering in a traditional Japanese dining room. The scene opens with a large, intricately patterned blue and white ceramic bowl of ramen in the immediate foreground, rendered in crisp, sharp focus. The bowl sits on a smooth, polished long wooden table. Inside the bowl, a rich, oily golden-brown broth surrounds yellow wavy noodles, topped with two thick, round slices of chashu pork featuring visible fat marbling and a distinct spiral meat pattern. A generous mound of freshly chopped, bright green scallions rests in the center, and a crisp, dark green rectangular sheet of nori seaweed is tucked into the right edge. To the left of the bowl, a pair of light brown wooden chopsticks rests horizontally on a small, dark rectangular chopstick rest, near a small cylindrical ceramic teacup with blue painted patterns. On the right side of the table, a spherical paper lantern with a ribbed bamboo frame sits on a black wooden base. In the background, a large family of seven is gathered around the table, initially appearing as a soft, blurred presence. Behind them, traditional Japanese sliding shoji screens with wooden lattice frames are open, revealing a bright outdoor scene with lush green trees. Early in the clip, the thick, white steam rising from the hot ramen broth immediately intensifies, billowing upwards in thick, swirling clouds that dance continuously above the bowl. As the clip progresses into the middle seconds, the camera maintains its static position while the focus begins a deliberate, smooth shift deeper into the room. The foreground ramen bowl, its vibrant ingredients, and the rising steam gradually soften into a hazy, out-of-focus blur. Simultaneously, the family members in the background come into sharp, detailed clarity. The heavy steam continues to rise from the foreground, creating a dynamic, translucent veil between the camera and the family. With the focus now firmly locked on the background, the vibrant family dinner comes alive. The man in the dark navy blue long-sleeved shirt on the left leans forward, his mouth moving animatedly in a silent exchange. The young girl in the crisp white short-sleeved t-shirt beside him smiles brightly, looking toward the center of the table. The woman on the far left, wearing a soft light blue long-sleeved blouse, turns her head slightly, smiling gently. Across the table, the woman in the light grey button-down shirt smiles broadly, her eyes crinkling, as she rests her hands near her plate. The woman in the dark grey top further back uses her wooden chopsticks to pick up a small piece of food from a central ceramic dish filled with bright red pickled vegetables. The woman in the center back in the light grey sweater smiles gently, her hands clasped softly in front of her, observing the interaction. Throughout the remainder of the clip, the family continues their lively physical interaction, their mouths moving in continuous, silent cadences of conversation, while the thick, white steam from the blurred ramen bowl in the foreground never stops rising, adding a comforting atmosphere to the warm gathering.\n\noverall_soundscape: The soundscape begins with a quiet room tone mixed with the faint, airy rustle of the thick steam billowing from the hot ramen bowl in the foreground, accompanied by the subtle, continuous hissing and bubbling of the rich broth. As the visual focus shifts deeper into the room, the physical sounds of the bustling family dinner become dominant in the foreground. The clear, sharp clinking of ceramic bowls and wooden chopsticks touching plates is clearly heard as the family members reach for food. This is followed by the faint, muffled thud of a cup being set down on the smooth wooden table, and the subtle, rhythmic rustle of cotton and wool clothing as the family members lean forward and gesture, perfectly capturing the lively, physical atmosphere of the shared meal.\n\nnon_diegetic_music: A gentle, heartwarming acoustic guitar melody plays softly in the background, accompanied by the subtle, resonant notes of a traditional Japanese koto. The music maintains a slow, comforting tempo that enhances the cozy, nostalgic, and joyful atmosphere of the family gathering." }, "duration": 8, "usage": { "total_tokens": 22822, "prompt_tokens": 12800, "completion_tokens": 10022 }, "ratio": "16:9", "task_type": "h3_context_ir", "modality": "text" } }
H3-Base	View script (opens in new tab)	i2va.mp4 (opens in new tab)
H3-Regenerate-2K	View script (opens in new tab)	i2va_2k.mp4 (opens in new tab)
Reference 2K result by directly calling Open Platform API	View script (opens in new tab)	i2va_direct_2k.mp4 (opens in new tab)
Reference 768P result by directly calling Open Platform API	View script (opens in new tab)	i2va_direct_768p.mp4 (opens in new tab)

case-Ref2VA

Type: Multimodal reference-to-video (video + audio)
Duration: 5 seconds
Aspect ratio: adaptive

stage	request	result
H3-Context-IR	View script (opens in new tab)	json { "task": { "id": "<task_id>", "model": "MiniMax-H3", "status": "succeeded", "created_at": "<created_at>", "updated_at": "<updated_at>", "content": { "prompt": "subject_definitions:\n<Subject 1> is the young man with short wavy blonde hair, wearing a bright pink suit jacket, matching pink trousers, an unbuttoned white shirt, and silver rings, holding a small black lamb in his arms in <Video 1>.\n<Video 1> is the source video for the editing task.\n<Audio 1> is the synchronized audio track of <Video 1>, providing the background music.\n<Audio 2> is the voice timbre reference for <Subject 1>'s voice, containing a spoken male voiceover.\n\nsummary:\n[video editing + audio reference + audio reuse] The target video is an edited version of <Video 1>. <Subject 1>, wearing a bright pink suit and holding a black lamb, stands in a grassy field with other white lambs in the background. The edit animates <Subject 1>'s face to speak the user-provided dialogue. <Audio 1> is partially reused as the continuous background music, while the target references the calm male voice timbre of <Audio 2> for <Subject 1>'s spoken lines.\n\nretention_analysis:\n<Subject 1> (appears in [Shot 1]): fully_preserved - the man retains his identity, wavy blonde hair, pink suit, white shirt, accessories, and the black lamb he holds, with his mouth newly animated to speak.\n<Video 1> (source video editing): fully_preserved - the original camera framing, warm golden hour lighting, grassy hill setting, and background white lambs are maintained while the central character is edited.\n<Audio 1>: partially_copy - the atmospheric background music from <Audio 1> is reused in the target video, mixed beneath the newly added spoken dialogue.\n<Audio 2>: reference - the target audio references the male voice timbre from <Audio 2> to generate <Subject 1>'s spoken dialogue.\n\ndetailed_description:\nThe target video is in realistic photographic style.\n[Shot 1] The shot begins from the source <Video 1>, showing <Subject 1>, a young man with short wavy blonde hair, wearing a bright pink suit jacket, matching pink trousers, and a casually unbuttoned white shirt. He stands confidently in a sunlit green pasture, gently holding a small black lamb securely in his arms. The warm, golden hour lighting casts soft shadows across his face and the bright pink fabric of his suit. Behind him, several white lambs stand and graze on the rolling grassy hill against a clear, pale blue sky. The atmospheric background music from <Audio 1> plays continuously throughout the scene. <Subject 1> physically speaks, his mouth movements naturally syncing to the new dialogue, with his voice timbre referencing the calm male delivery from <Audio 2>. Looking thoughtfully forward, <Subject 1> (S1) speaks softly, <d>[English] Follow the wind, live free.</d> As he delivers the line, he subtly shifts his weight, cradling the resting black lamb while the camera slowly pushes in. <Subject 1> (S1) continues his thought, <d>[English] Leave worries behind, enjoy the moment.</d> Exactly as his voice stops, his lips meet in a relaxed, peaceful smile, and his jaw ceases speaking motion. He then turns his gaze slightly away toward the horizon, gently stroking the black lamb's fleece with his fingers as the camera holds on this tranquil, sunlit state through the end of the video.\n\noverall_soundscape:\nThe soundscape consists of the continuous, atmospheric background music from <Audio 1>, overlaid with the clear, calm male dialogue spoken by the main character, referencing the voice timbre of <Audio 2>.\n\nnon_diegetic_music:\nThe atmospheric, sustained background music from <Audio 1> is reused as the continuous score, playing quietly beneath the spoken dialogue." }, "duration": 5, "usage": { "total_tokens": 39299, "prompt_tokens": 33323, "completion_tokens": 5976 }, "ratio": "16:9", "task_type": "h3_context_ir", "modality": "text" } }
H3-Base	View script (opens in new tab)	r2va.mp4 (opens in new tab)
Reference 2K result by directly calling Open Platform API	View script (opens in new tab)	r2va_2k.mp4 (opens in new tab)
H3 API 2K in Open Platform for reference	View script (opens in new tab)	r2va_direct_2k.mp4 (opens in new tab)
Reference 768P result by directly calling Open Platform API	View script (opens in new tab)	r2va_direct_768p.mp4 (opens in new tab)

Prompting Guidance

VIDEO_PROMPT_WRITING_GUIDE_base_en.md (opens in new tab)

VIDEO_PROMPT_WRITING_GUIDE_ref_en.md (opens in new tab)

License

MiniMax H3 is released under the MiniMax H3 Community License Agreement (opens in new tab). Q&A about the License (opens in new tab)

Contact Us

Safetensors

Model size

33B params

Tensor type

F32 BF16

Model tree for MiniMaxAI/MiniMax-H3

Quantizations

2 models (opens in new tab)

Spaces using MiniMaxAI/MiniMax-H3 7

Collection including MiniMaxAI/MiniMax-H32 items • Updated • 3 (opens in new tab)

Insights

Research extraction

1. Model size and precision

H3-Omni-Transformer is a 33B-parameter dense, single-stream Transformer; ~13B parameters are in AdaLN branches whose outputs can be precomputed/cached, so they are not needed for inference-only deployment; full weights are released for fine-tuning.
Card metadata lists "33B params" and tensor types "F32 BF16".

2. Architecture and capabilities

System = H3-Context-IR (hosted preprocessing/orchestration, critical for output quality, not in this open-source release; API provided) → H3-Base (768p video+audio generation) → H3-Regenerate-2K (2K in-context regeneration, not yet open-sourced; API provided).
Output: 4–15 s video at 24 FPS, default 768p (2K via regeneration), 32 kHz stereo audio, stable dialogue support in 11 languages.
H3-Base: text via H3-Encoder, visuals via H3-Encoder + H3-VisualVAE, audio via H3-AudioVAE, packed into one sequence with RoPE; H3-Omni-Transformer jointly predicts video and audio latents.
H3-Encoder uses full pretrained Qwen3-VL-32B weights (hidden states from layer 50); the H3 tokenizer config with special tokens such as is required.
H3-VisualVAE: spatial 16×, temporal 4×, 24 channels (f16t4d24); tokens patchified 1×2×2 → effective spatial 32×.
H3-AudioVAE compresses 32 kHz audio to 40 Hz latent tokens per channel (shared encoder/decoder across stereo channels).
Attention and FFN have no modality-specific structure; 3D MM-RoPE encodes (t,h,w).
Sparse attention is natively supported but absent from the initial release (inference is full-attention only).

3. Reasoning / tool-use

No tool-use or agentic capabilities are described. The only reasoning reference is internal to H3-Context-IR, whose workflow includes "complex logical reasoning" for interpreting multimodal instructions.

4. Context length — gap

Not specified in the card. Documented limits are output duration (4–15 s), reference-input caps (≤9 images, ≤3 video clips 2–15 s, ≤3 audio clips, max 12 files total), and API token-usage examples (e.g., 39,299 total tokens for a 5 s Ref2VA task); none define the H3 context window.

5. Licensing

Released under the MiniMax H3 Community License Agreement (Q&A doc linked); guardrails do not affect licensee obligations on lawful use and use restrictions.

6. Local inference / setup

Two BF16 CFG-distilled task checkpoints: H3-Base FL2VA (t2va/fl2va) and H3-Base Ref2VA (ref2va), each with processor, tokenizer, text encoder, transformer, VisualVAE, AudioVAE.
Repo hosts original checkpoints and diffusers format side by side: hf download MiniMaxAI/MiniMax-H3 --include "FL2VA/*" "Ref2VA/*" for SGLang/vLLM; diffusers can use ModularPipeline.from_pretrained("MiniMaxAI/MiniMax-H3").
Recommended frameworks: SGLang, vLLM, diffusers, ComfyUI (T2V/R2V templates).
SGLang reference: sglang serve --model-path MiniMaxAI/MiniMax-H3 --num-gpus 4 --ulysses-degree 4 --performance-mode speed --model-variant fl2va|ref2va.
Full 2K reproduction combines local SGLang with hosted H3-Context-IR and H3-Regenerate-2K APIs.

7. Coding benchmarks / pricing — gaps

No coding benchmarks or prices appear anywhere in the card; it only links API docs (platform.minimax.io / platform.minimaxi.com) and apps (hailuoai.video, hub.minimax.io).

Hugging Face