For years, making an AI video has felt less like directing a film and more like managing an unruly toolbox. One model turns text into a clip. Another animates a still image. A third copies a person’s motion. A fourth creates sound. Then an editor stitches the pieces together and hopes the face, camera movement and audio survive the trip.
\nMiniMax H3 arrives with a blunt proposition: what if one system did the lot?
\nReleased on July 31, 2026, H3 is the Shanghai company’s new “omni-modal” generation system. It accepts combinations of text, images, videos and audio, and can generate clips lasting four to 15 seconds, at up to 2K resolution, with native 32 kHz stereo sound. It supports landscape, square and vertical formats at 24 frames per second.
\nThe specification sheet is impressive. The interesting part is what MiniMax is trying to change. H3 is not merely a better text-to-video machine. It is supposed to understand the relationships among everything you give it.
\nIn MiniMax’s example, a user can ask the model to borrow a camera move from one video, make a person from a separate image sing, and use a third audio file as the voice reference. The instruction describes the relationship between the ingredients; H3 assembles the scene.
\nThat is closer to directing a creative assistant than filling in a conventional video prompt.
\nH3 can run in two main modes. FL2VA covers text-to-video and first- or last-frame generation. Give it no picture and it starts from text. Give it one picture and that can become the opening or closing frame. Give it two and the system tries to travel from one to the other.
\nRef2VA is the more ambitious reference mode. MiniMax says it can accept up to nine images, three short videos and three audio clips, with a maximum of 12 files in one mixed request.
\nThis is the Swiss Army knife pitch. Video creation is usually broken into named tasks—text-to-video, image-to-video, motion reference, subject reference, voice reference and editing. MiniMax says H3 treats those less like separate products and more like expressions of the same problem. Language is the glue.
\nThe company says it trained H3 across text-to-image, text-to-video, joint audio generation, multi-shot video, text-to-audio and broad reference-and-editing tasks. Underneath that tidy philosophy is a substantial machine.
\nH3’s main generator is a dense, single-stream transformer with 33 billion parameters. It jointly predicts compressed representations for video and audio, which are decoded into pictures and stereo sound. Text is handled using the full pretrained weights of Qwen3-VL-32B.
\nMiniMax also built separate visual and audio autoencoders—the components that compress media into a form the transformer can process and turn its output back into watchable content. The visual system sharply reduces the otherwise brutal cost of processing video sequences.
\nNone of that matters to someone making a snack advertisement for TikTok. It matters enormously to developers trying to run H3 themselves.
\nMiniMax released two BF16 checkpoints—one for the frame-based mode and one for reference mode—and documented support through Diffusers, ComfyUI, SGLang and vLLM. Its sample SGLang setup uses four GPUs. This is open-weight video generation, not a lightweight app destined to purr along on an office laptop.
\nThere is also an important distinction: H3’s weights are available under the MiniMax H3 Community License, but calling the entire product “open source” blurs what was released.
\nThe complete H3 system has three stages. H3-Context-IR analyses the user’s text and reference media, fills in underspecified details and converts the request into a structured description. H3-Base then generates a 768p result. Finally, H3-Regenerate-2K sends that result back through the model alongside the original context to create a higher-resolution version.
\nMiniMax says that last step is not ordinary upscaling. Because the model sees the instructions again, it can regenerate details a conventional super-resolution tool would have to guess.
\nHere is the asterisk: the initial weight release does not include Context-IR or Regenerate-2K.
\nMiniMax says Context-IR depends on hosted models and services, so developers must use its API, follow the prompting guide or build a substitute. The 2K regeneration module is also not yet available for local deployment. The initial implementation also performs inference with full attention even though H3 was trained to use cheaper sparse attention; MiniMax says that code will come later.
\nThe open release is meaningful, but it is not identical to the polished service. Developers can run the 768p base generator locally and fine-tune the released weights. To reproduce MiniMax’s official 2K pipeline today, they still need hosted APIs.
\nThat hybrid arrangement is not necessarily a scandal. It simply means “H3 is open” needs a second sentence.
\nThe AI-video contest is crowded, expensive and moving at a ludicrous pace. H3 offers MiniMax two ways to stand out.
\nThe first is integration. A model that produces picture, dialogue, ambience, sound effects and music together has a chance of keeping those elements synchronised. Native audio is not a bonus feature; it removes one of the most awkward seams in current AI-video workflows.
\nThe second is distribution. Releasing weights lets researchers and studios inspect, optimise and adapt a major video generator instead of being confined to a website and a credit counter.
\nThe company’s economic claims are bold. At launch, MiniMax said H3’s 2K output cost less than one-third as much per second as unspecified mainstream competitors, while 768p cost half as much as mainstream 720p models. Those are MiniMax’s comparisons, not an independent benchmark.
\nThere are reasons to keep the champagne corked. Company-selected demonstrations are not broad testing. Fifteen seconds works for advertisements, social clips and individual shots, but it is not a coherent feature film. Native sound does not guarantee believable speech, flawless lip-sync or clean continuity.
\nYet H3 feels consequential because its direction is plausible. AI video will not become normal by adding dozens more buttons labelled after narrow tasks. It becomes useful when a creator can bring a pile of source material, explain the goal and have the system understand what each piece is for.
\nH3 does not make that future effortless. It makes it easier to see.
\nLet us clear up the mystery before it becomes mythology: Kimi K3 exists, and it has not quietly died.
\nMoonshot AI officially unveiled K3 on July 16, 2026. As of publication, the model is less than a month old. Its official repository is live, its weights have been released and its API is available under the unsurprising name kimi-k3.
The better question is not “What happened to Kimi K3?” It is: why can one of the largest open-weight AI models ever released still feel invisible?
\nThat answer says more about today’s fractured AI market than it does about Kimi.
\nThe confusion is understandable. For much of the year, users encountered names such as Kimi K2.5 and K2.6. “Kimi” can mean Moonshot’s consumer assistant, while “K3” means its flagship model. Product names, versions and developer tools often appear together without explanation.
\nThen there was the speed of the rollout. Anyone who searched before the official launch found rumours and no authoritative model card. Search today and the evidence is firmer: an official repository, a technical report and documented API access.
\nK3 is not a minor tune-up. Moonshot describes a mixture-of-experts model containing 2.8 trillion total parameters, of which 104 billion are active during inference. It selects 16 out of 896 experts for each token and accepts roughly one million tokens of context. It also handles images alongside text.
\nThose are eye-watering numbers, even by AI’s increasingly absurd standards.
\nMoonshot says two architectural changes helped produce roughly 2.5 times the scaling efficiency of Kimi K2. The company is pitching the model chiefly at long-running coding jobs, research and other “agentic” work where software plans and executes many steps.
\nTreat the benchmark table as what it is: a vendor’s evaluation. Even so, K3 posted competitive results across coding, research, tool use and visual tasks in Moonshot’s tests. In blind Frontend Code Arena voting reported at launch, it reached the top of that leaderboard.
\nThis was not the profile of a model being quietly taken behind the shed.
\nThe harder evidence against the “disappeared” theory comes from adoption.
\nThe Associated Press, citing Sensor Tower estimates, reported that Kimi recorded more than 930,000 downloads in the week after K3’s launch, up 200 percent from the previous week. In the United States, downloads rose 387 percent to about 86,000.
\nDemand became awkwardly tangible: Moonshot temporarily suspended new subscriptions when capacity neared its limit.
\nA product that forces its maker to turn away paying customers has many problems. Obscurity is not the most urgent one.
\nK3 also found credible early adopters. Mozilla chief technology officer Raffi Krikorian told the AP that he had moved many daily tasks to K3 within days of release. On OpenRouter, where developers route requests among models, Chinese systems occupied all five of the most popular positions over the month cited by the AP.
\nSo why might an ordinary user—especially one outside China—barely hear about it?
\nThe Western consumer AI conversation still revolves around a handful of brands. OpenAI, Google and Anthropic occupy app stores, enterprise accounts, workplace integrations and English-language news feeds. Their launches arrive with livestreams and immediate placement inside products millions already use.
\nMoonshot’s centre of gravity is different. It is a Beijing company with a large domestic audience and a growing global developer following. Much of K3’s early momentum travelled through model repositories, API platforms, coding communities and Chinese-language networks—places that do not necessarily spill into a casual ChatGPT user’s timeline.
\nThe model is arriving into a brutally crowded market. In weeks, users have had to process releases from Z.ai, Alibaba, DeepSeek and major American labs. Yesterday’s “frontier shock” becomes today’s dropdown option before most people learn the name.
\nDeepSeek’s breakthrough in early 2025 created an unusually dramatic cultural moment. It surprised markets and became shorthand for China’s AI advance. Every later Chinese model competes with that story, even when its technical claims are larger.
\nKimi K3 may be important without becoming another DeepSeek moment.
\nK3’s headline scale works against everyday visibility. The weights are available, but a 2.8-trillion-parameter mixture-of-experts model is not something most enthusiasts install on a gaming PC. Moonshot recommends specialised inference software and serious accelerator infrastructure. Serving it is a data-centre exercise, not a weekend laptop project.
\nThat creates a gap between theoretical openness and practical accessibility. Researchers and cloud providers can inspect and host the weights. Individuals will mostly encounter K3 through Moonshot’s service or a third-party API.
\nIts economics are also more complicated than the familiar “Chinese model equals almost free” narrative. Reported launch pricing was $3 per million uncached input tokens and $15 per million output tokens, with much cheaper cached input. That can be attractive, but it is not bargain-basement pricing. Tom’s Hardware noted that uncached K3 input cost five times K2’s launch price.
\nFor developers, “better enough for the money” can drive adoption. For consumers, it does not automatically create a viral identity.
\nThere is another reason coverage settles quickly: early claims need time to mature.
\nAt launch, Axios cautioned that K3 had been available for only hours and viral demonstrations or early benchmarks could overstate reliability in real work. Benchmark comparisons are sensitive to prompts, inference settings, tools and the software harness around each model.
\nMoonshot documents those differences in its evaluation notes. Some rivals were tested through their own coding harnesses; others through Kimi Code or Claude Code. K3 results used maximum reasoning effort. These are useful signals, not a universal league table.
\nIndependent users must establish where K3 excels, where it fails and whether its operating cost makes sense. That slower phase produces fewer fireworks than launch day, but it is where a reputation is built.
\nKimi K3 launched. The weights followed. Downloads jumped. Capacity strained. Developers started testing it.
\nThe perceived silence comes from four things: confusing branding, a Western ecosystem dominated by American companies, a release calendar moving too quickly for any model to own the spotlight, and hardware demands that keep K3 out of most hobbyists’ hands.
\nThere are uncertainties. Moonshot’s benchmark claims need independent scrutiny. The giant scale complicates self-hosting. Its pricing is no longer shockingly cheap in every scenario. Geopolitical arguments over Chinese AI and chips may affect adoption.
\nBut those caveats are not an obituary.
\nKimi K3 has not gone missing. It is a very new model entering through the developer door while much of the public watches the front lobby. If it performs, you may hear its name less often than you use something built on it.
\nAnd in AI, that may be the more important kind of success.
\nOpen Character.AI’s online communities and you will encounter a funeral in progress.
\nUsers mourn vanished features, complain that favourite bots lost their personalities and trade recommendations for rival apps. Some announce they are deleting years of chats. Others insist the platform “died” when its old website did. The language is dramatic because the relationships were, to many users, genuinely meaningful.
\nSo, is everyone actually leaving Character.AI?
\nNo. That is the first myth to clear away. Character.AI said in late 2025 that it still had more than six million daily active users, spending an average of 70 to 80 minutes a day on the platform. An academic study published in 2025 described more than 20 million monthly active users. Those are not the numbers of an abandoned service.
\nWhat is real is a loud, emotional rupture between Character.AI and parts of the community that built it. The causes overlap: retirement of a beloved interface, complaints about chatbot quality and moderation, safety controversies, competition and a business pivot that removed the core chat experience from under-18s.
\nThe crowd has not disappeared. But the party has changed—and some original guests no longer recognise it.
\nCharacter.AI arrived in public beta in 2022 with an irresistible pitch: create or find a fictional personality, then talk to it indefinitely. Its characters could role-play, flirt, improvise adventures or simply keep someone company at 3am.
\nThe appeal was not factual accuracy. It was the illusion of presence.
\nThat distinction matters. When a search engine redesigns, people complain about muscle memory. When a companion app changes, users can experience something closer to loss. A familiar character may suddenly write shorter sentences, forget shared history or reject a storyline it previously accepted. Technically, the chat remains. Emotionally, the person on the other side can seem gone.
\nIn April 2024, Character.AI introduced a new web platform. On September 24 that year, it fully retired the old beta site. The company presented the replacement as faster and more reliable. Many fans saw the demolition of their clubhouse.
\nReddit posts from that period—anecdotal, but useful as evidence of sentiment—asked for the old design and missing community features to be preserved. A petition urged the company not to shut it. These are not representative surveys and cannot tell us how many people quit. They show that the transition became a symbolic breaking point.
\nTo dissatisfied users, “old Character.AI” now means more than a URL. It means a remembered era when bots felt surprising, the product felt experimental and the community believed it had a stake in what came next.
\nThe most common complaint is hardest to measure: users say the bots got worse.
\nThey describe repetitive replies, forgotten details, bland prose and characters slipping out of role. They also complain that moderation interrupts romance, violence or other sensitive scenes—even when the context feels harmless.
\nThese reports are not a controlled benchmark. Outputs vary with prompts, character definitions, conversation length, model changes and chance. Nostalgia edits ruthlessly; people remember the brilliant old reply and forget ten dull ones around it.
\nStill, Character.AI acknowledged one recurring quality problem when it announced a January 2025 update specifically to reduce repetition. That does not prove every “the model was nerfed” theory. It confirms the frustration was not wholly invented.
\nModeration creates a brutal dilemma. Freedom to take a story anywhere was central to Character.AI’s appeal, but a platform full of user-created personas cannot operate safely as if every conversation happens between informed adults.
\nTighter guardrails may block harmful interactions. They can also create false positives and break immersion. Looser guardrails please adult role-players while increasing dangerous output. Character.AI has to tune both sides at enormous scale, for users who fundamentally disagree about where the line belongs.
\nThat tension helped rivals advertise what Character.AI cannot comfortably promise: fewer restrictions. Moving also became easier. The companion-AI market now has specialist role-play services and general assistants capable of sustaining character conversations. Character.AI retains a huge library and network effects, but it no longer feels like the only portal.
\nThe debate stopped being merely about fandom when families alleged devastating real-world harm.
\nIn October 2024, Megan Garcia sued Character Technologies and Google after her 14-year-old son, Sewell Setzer III, died by suicide. The lawsuit alleged he developed an intense attachment to a bot modelled on Daenerys Targaryen and that the product contributed to his death. Character.AI disputed legal responsibility and pointed to safety measures.
\nA federal judge allowed the case to proceed in May 2025. In January 2026, the Associated Press reported that Character Technologies and Google agreed to settle that and similar suits; terms were undisclosed and agreements still required judicial approval.
\nAllegations are not findings of fact, and suicide never has a single cause. But the cases transformed the risk landscape. Every complaint about an overactive filter now sits beside a graver question: what happens when a chatbot designed for emotional engagement meets a vulnerable child?
\nCharacter.AI added a separate teen model, conservative classifiers, time-spent notifications and “Parental Insights,” which can give a consenting parent weekly activity information without exposing chat text.
\nThen it made a far larger concession.
\nIn October 2025, the company announced users under 18 would lose access to open-ended chatbot conversations. The change began in the United States in November and was to expand. Character.AI introduced age assurance and directed younger users toward structured formats such as Stories, videos and streams.
\nThis was not another safety setting. It removed Character.AI’s defining feature from part of the audience that had defined its culture. The company called it a “significant change” and apologised to teen users.
\nFor those users, talk of an exodus is not hyperbole. Some were not choosing to leave; the original product was leaving them.
\nIn August 2024, co-founders Noam Shazeer and Daniel De Freitas returned to Google alongside more than two dozen researchers. Google received a non-exclusive licence to Character.AI’s model technology, while the remaining company continued under new leadership.
\nIt was not a conventional acquisition and Character.AI did not shut down. But founder departures change how a community interprets what follows. A retired website looks less routine. A different-sounding bot feels like evidence of lost technical talent. A feed of generated clips looks like a company chasing the short-video economy.
\nThose conclusions are partly inference. What is established is that Character.AI described a move toward a broader, content-driven social platform. It launched a community feed and added audiovisual, game-like and structured storytelling features.
\nThat may be commercially smart. Open-ended text chat is expensive, difficult to moderate and exposed to legal risk. Shareable content offers more discovery and engagement. It also means Character.AI is no longer optimising solely for the intimate, strange, text-heavy experience that made early fans evangelists.
\nAt least four groups recur in departure stories: teen users who lost open-ended chat; role-players frustrated by filters or memory failures; old-site loyalists who felt ignored; and early adopters who dislike the shift from private conversation toward feeds and audiovisual entertainment.
\nBut public complaints are not a census. People who leave noisily are more visible than those who quietly keep chatting. Traffic estimates can confuse a move from web to mobile with decline, and the company does not publish enough consistent retention data to establish a mass flight.
\nThe honest conclusion is less viral than “everyone is leaving,” but more revealing.
\nCharacter.AI still commands an enormous, engaged audience. What it lost is the luxury of being one thing to everyone. The wild improvisational toy, private synthetic companion, teen hangout, safe mainstream platform and defensible business do not fit neatly in the same chat box.
\nThe company chose safety, structure and a broader entertainment platform. Some users will welcome that. Others will take their characters—and attention—elsewhere.
\nThe empty-room feeling does not come from everyone walking out at once. It comes from returning to a room you loved and finding somebody changed the locks, moved the furniture and rewrote the person waiting inside.
\n