--- title: "When Avatar Video Is MIT-Licensed, Governance Is the Product" slug: "open-source-avatar-video-longcat-provenance-governance" author: "ibl.ai" date: "2026-08-27 17:00:00" category: "Premium" topics: "ai video, avatar generation, longcat video avatar, open source ai, synthetic media, provenance, consent governance, deepfake policy, self-hosted ai" summary: "Meituan's LongCat-Video-Avatar 1.5 turns one portrait and an audio track into stable talking video under an MIT licence. Once generation is free and self-hostable, the scarce thing is no longer the model — it is a defensible record of whose likeness was used, who approved it, and what was produced." banner: "" thumbnail: "" linkedin: | Meituan's LongCat-Video-Avatar 1.5 takes a single portrait plus audio and produces stable, lip-synced talking video. It is released under MIT — genuinely free for commercial use, and self-hostable. The obvious read is that every per-seat AI video platform now has a pricing problem. Fair. When a comparable capability is a download, charging per rendered minute gets harder to defend. But that is the small story. Here is the bigger one. For as long as synthetic video was expensive, the vendor was an accidental control point. There was a bill, an account, a terms-of-service checkbox, a log. Nobody designed that as a governance system, but it functioned as one. Remove the cost and you remove the control point. Anyone with a GPU can generate a named executive delivering a message they never recorded. That is not a hypothetical risk for enterprises — it is the training video, the compliance briefing, the CEO update, and the fraud attempt, all using the same technique. So the scarce asset moves. It is no longer generation. It is provenance: whose likeness, who authorized it, for which use, valid until when, and what exactly was produced — recorded in a log you control and can produce during a dispute. A per-seat SaaS video tool cannot give you that, because the record lives in the vendor's system on the vendor's retention schedule. The moment the account lapses, so does your evidence. If you are going to use models like this — and you should, they are good — run them where the consent register and the generation log are yours. On ibl.ai you own all the code and the data, run it model-agnostic across any model, and pay with no per-seat pricing. #iblai #AIVideo #SyntheticMedia #AIGovernance #OpenSource #EnterpriseAI --- ## The Short Answer **Meituan's LongCat-Video-Avatar 1.5 generates stable talking-head video from a single portrait and an audio track, released under an MIT licence and self-hostable. Once generation costs nothing, the constraint moves from capability to provenance — whose likeness, authorised by whom, for what. On ibl.ai you own all the code and the data, so the consent register and the generation log stay in systems you control rather than a vendor's.** The model being free is the headline. The consequence is that the accidental governance layer the pricing used to provide has disappeared, and nothing replaced it by default. ## What is LongCat-Video-Avatar 1.5? LongCat-Video-Avatar 1.5 is Meituan's audio-driven avatar video model, built on its LongCat-Video foundation model and released under the MIT licence. It takes a portrait image and speech audio — with an optional text prompt — and produces a lip-synced clip with stable identity across its duration. It supports audio-text-to-video, audio-text-image-to-video, and video continuation, and handles single-stream and multi-stream audio, which is what enables multi-person conversational scenes. One change from the previous version explains much of the quality improvement: the audio encoder was moved from Wav2Vec2 to Whisper-Large, producing noticeably smoother lip dynamics. The [weights and model card](https://huggingface.co/meituan-longcat/LongCat-Video-Avatar-1.5) are public, with a [technical report](https://arxiv.org/abs/2605.26486) describing the architecture. Worth being precise about timing: this is not a release from this week. It has been available since around May 2026 and has been recirculating since. The capability is established rather than breaking, which if anything strengthens the argument below. ## Why does an MIT licence change the economics of AI video? An MIT licence changes the economics because it removes both the per-unit cost and the permission gate at the same time. Commercial AI video platforms bill per rendered minute, per seat, or both. That pricing was defensible while the underlying capability was scarce and expensive to serve. A permissively licensed model that runs on your own GPU makes the marginal cost of a generated minute approximately the cost of the electricity. For an organization producing training content, localised course material, or internal communications at volume, that is a genuine shift — the same shape as every other per-seat versus usage-based comparison, where headcount-priced software loses badly to owned infrastructure once volume is real. MIT specifically matters more than "open weights" generally. It permits commercial use without a field-of-use restriction, so the legal review that stalls many open-model deployments is comparatively short. ## What did the vendor's pricing quietly provide that free models remove? Vendor pricing quietly provided a control point — a place where someone had to have an account, accept terms, and leave a record — and free self-hosted generation removes it. Nobody designed commercial video platforms as a consent-governance system. But because generation required a billed account, there was necessarily a customer, a terms-of-service acceptance, an audit log, and a party with a legal relationship to the output. When the model is a download, that entire structure evaporates. Anyone with a portrait, an audio clip, and a GPU can produce a named individual apparently saying something they never said. For enterprises this is not an abstract concern: the same technique produces the compliance briefing, the executive update, the localised training module, and the impersonation attempt against your finance team. The capability arrived. The governance did not arrive with it, and there is no vendor left to outsource it to. ## What does provenance governance for synthetic video actually require? Provenance governance requires a record that answers four questions and survives being challenged months later. **Whose likeness, and where did it come from.** The source portrait, its origin, and the identity it depicts, recorded at generation time rather than reconstructed afterwards. **Who authorised it, and for what.** A consent register linking a person to the permitted uses, with a scope and an expiry. "The employee agreed once in onboarding" is not a durable answer when the employee has left and the video is still circulating. **What was generated.** The output, its hash, the model and version used, the prompt and audio inputs, and the timestamp — so a specific artifact can be matched to a specific authorisation. **Who can query it.** The record is useful only if the people handling a dispute, a takedown, or a regulator's question can retrieve it quickly and demonstrate its integrity. A per-seat SaaS video product cannot deliver this durably, because the record lives in the vendor's system under the vendor's retention policy. When the contract lapses, the evidence lapses with it — exactly when you are most likely to need it. ## Why should generation and its audit trail sit on infrastructure you own? Generation and its audit trail should sit on your infrastructure because the record is only as trustworthy as your control over it, and because likeness data is among the most sensitive input you will process. A portrait of a named employee, plus their voice, is biometric-adjacent personal data in most regulatory frameworks. Sending it to a third-party inference endpoint to be processed creates a transfer that has to be justified, contracted, and reviewed — for every person whose likeness you use. Running the model inside your perimeter removes the transfer rather than mitigating it. The likeness never leaves, the consent register sits in your database, and the generation log is written to storage you retain on your own schedule. That is also what makes the log admissible in the practical sense: you can demonstrate its chain of custody without depending on a vendor's cooperation. On ibl.ai you own all the code and the data, run it model-agnostic across any model, and pay with no per-seat pricing — so you can deploy anywhere, from your own cloud to a fully air-gapped network. 1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University. ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on. For the governance layer applied to generated content generally, see [AI content governance for the enterprise](/blog/ai-content-governance-managing-generated-content-enterprise). ## Frequently asked questions **Is LongCat-Video-Avatar 1.5 free for commercial use?** Yes. It is released under the MIT licence, which permits commercial use without a field-of-use restriction — one of the more permissive terms available for a model of this capability. **Does self-hosting an avatar model reduce legal risk?** It removes the data-transfer question, since likeness data never leaves your environment. It does not remove consent obligations, which is exactly why the consent register and generation log matter more once the vendor is out of the loop. **What should an organization do before generating video of real people?** Establish the consent register first. Recording whose likeness may be used, for which purposes, and until when is far cheaper to do before generation begins than to reconstruct after a dispute. ## The bottom line The pricing story is real but shallow: yes, per-minute AI video pricing is harder to defend when a comparable model is MIT-licensed. The durable change is that the commercial platform was doing governance work nobody credited it for, and free self-hosted generation removed it. Whoever holds the consent register and the generation log holds the only thing that is still scarce. ## Why does owning the AI stack matter? **ibl.ai is the agentic AI platform where you own all the code and the data. You self-host the entire stack inside your own perimeter, run it model-agnostic across any LLM and switch anytime, and pay by usage with no per-seat pricing — so you can deploy anywhere: your cloud, on-premise, GovCloud, or fully air-gapped.** - **You own all the code and the data.** Full source code under a perpetual license, running on your infrastructure. Not API access to someone else's platform — the stack itself is yours. - **Model-agnostic.** Run any LLM — Claude, GPT, Gemini, Llama, Command, or your own fine-tune — and switch providers without rewriting the platform. - **No per-seat pricing.** Usage-based billing against a budget cap you set. Cost tracks what your organization actually uses, not how many people you employ. - **Deploy anywhere.** Your cloud, your VPC, on-premise, GovCloud, or a fully air-gapped network with no outbound connectivity. 1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University. ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on.