Atmanity, an agentic AI lab founded in 2024, has launched Atmee v2, a real-time video model for live avatars, and Causeur, a model that controls how the avatar behaves. Per the company, one image and a short audio clip are enough to create a character you can talk to live, on a single GPU.
Atmanity also claims Atmee v2 is up to 50 times faster and 10 to 100 times cheaper to operate than character models such as Wan 3.0 and Seedance 2.5. Those figures come from the company as relayed in launch coverage; the site itself does not state them. This post separates what the Atmanity site states from what you can check today.
TL;DR: what is claimed, what is checked
| Question | Answer |
|---|---|
| What launched? | Atmee v2 (video model) and Causeur (behavior model) |
| Input? | One image plus 15 seconds of audio per Atmanity's site; press said 10 seconds |
| Hardware? | Live rendering on a single GPU, per Atmanity |
| Speed and cost? | Up to 50x faster and 10 to 100x cheaper than Wan 3.0 and Seedance 2.5 (company claim) |
| Access? | Invitation-only API; request a code by email |
| Independent test? | None found |
| Main use cases? | Virtual tutors, sales training, interactive game characters, per launch coverage |
What do Atmee v2 and Causeur do?
Atmanity splits a live video agent into two parts. In its own words, "Causeur decides, Atmee v2 renders."
- Causeur is the behavior model. Atmanity says it handles "personality, emotion, timing and active listening, moment to moment." It decides what the character does while you speak and while it speaks.
- Atmee v2 is the renderer: "a real-time, full-frame video foundation model for live avatars. Face, body, micro-expression, emotion, and the environment."
The company's pitch is that video agents should behave like someone, "not something." That is a marketing line, but it points to a real design question. A talking head that moves its lips is easy. A character that nods at the right moment, reacts to your tone and holds a consistent personality is much harder.
Atmanity lists four features: real-time rendering on one GPU, any form or style (photoreal, stylized or non-human), distinct personalities, and a data engine fed by conversations on its consumer app, atmee.ai. The company says that app contributes thousands of hours of real conversations every month.
Who is behind it?
The site says Atmanity was founded in October 2024 by a team of AI PhDs whose previous company, Phiar, was acquired by Google in 2022. CEO Chen-Ping Yu founded Phiar. The team lists six PhDs, over 100 publications and a set of related papers, including work on listening-head generation, personalized avatars and multimodal conversational behavior. That background is verifiable on the site and in the publications it lists. It does not validate the new speed and cost numbers.
Is it 50x faster and 10 to 100x cheaper?
This is the central claim, and it deserves care.
What the claim compares. Wan 3.0 and Seedance 2.5 are general-purpose video generators. We covered Alibaba's Wan 3.0 and ByteDance's Seedance 2.5 at launch. They create clips with rich scenes, camera moves and long durations. Atmee v2 aims at one thing: a live talking character. A specialized model that does one narrow job on one GPU will usually beat a general model on cost and speed for that job.
What is missing. We did not find the resolution, frame rate, latency to first frame, number of concurrent streams per GPU, or which GPU. Without these, "50x faster" has no fixed meaning. Cost depends on whether you count GPU time, idle capacity and API markup.
What would verify it. A third-party test that feeds the same image and audio to each system, measures end-to-end latency and frames per second, and reports hardware and settings. Until then, read the numbers as a vendor's direction of travel.
How does it compare with other avatar systems?
Real-time avatars are an active area. For context, see our coverage of HeyGen's LiveAvatar and the open-source demos around GPT Live, the open-source LongCat talking avatar model, and how McKinsey used AI avatars in client job interviews. Each shows a different pressure: product quality, open weights, and real-world use.
| System | Type | Access | Our coverage |
|---|---|---|---|
| Atmee v2 and Causeur | Live avatar plus behavior model | Invitation-only API | This post |
| HeyGen LiveAvatar | Commercial live avatar | Product and API | HeyGen post |
| LongCat avatar | Open-source talking avatar | Open | LongCat post |
| Wan 3.0 | General video model | Released model | Wan 3.0 post |
| Seedance 2.5 | General video model | ByteDance product | Seedance 2.5 post |
How can developers try it?
Access is gated. The site says to request an Atmee v2 invitation code by email for API access, design partnerships or research collaboration. Atmanity's LiveKit plugin, version 0.2.0 on PyPI (released November 3, 2026), can already render both v1 and v2 avatars. Its README marks v2 as invite-only and says the first v2 session after a quiet period can take a few minutes to start. The plugin lets the avatar be dropped into a real-time voice agent stack.
If you get access, test these first:
- Time to first frame. Measure how long from the end of the user's sentence to the first moving frame.
- Lip-sync under noise. Try background noise, accents and fast speech.
- Idle behavior. Watch the character for 60 seconds of silence. Loops, freezing and uncanny blinking show up quickly.
- Identity drift. Run a 20-minute session and compare the face at the start and end.
- Cost per hour. Ask for a written price at your concurrency level.
- Latency under load. Run several streams at once on the same GPU.
Why live video agents are the next step after voice
Atmanity frames its lab around a simple trend: "AI agents went from text to voice. Video is next." Voice agents already handle phone support and tutoring. A face adds presence, which matters for coaching, language practice and sales role-play, where eye contact and reaction timing change how people behave.
The technical barrier is latency. A voice agent can stream audio with a small delay. A video agent must render full frames, often at 24 or 30 frames per second, while staying in sync with speech and while reacting to what the user is saying. That is why a one-GPU live design is a notable claim. Running many streams per GPU is what turns a demo into a business. It is also why the Atmee v2 integration with real-time stacks such as LiveKit's plugin on PyPI matters: developers already building voice agents can add a face without rebuilding the pipeline.
The early use cases named in launch coverage are virtual tutors, sales training and interactive game characters. Each has a different tolerance for error. A game character can be odd and still charming. A sales trainer that freezes mid-sentence undermines the lesson. A tutor for children needs strong safety controls. Match your pilot to the use case with the most forgiving quality bar.
What readers are asking
Do I need a GPU to use it? Atmanity says rendering runs on a single GPU on its side. The invitation-only API suggests you call a hosted service rather than run the model yourself.
Is it open source? Not as far as we can tell. The site offers invitation codes and partnerships, not weights.
Can it make non-human characters? The site says "any form, any style," including photoreal, stylized and non-human looks. Test unusual designs, because models trained mostly on human faces often struggle there.
How is this different from a video call filter? A filter maps your own face onto a character. Atmee v2 generates the character from an image and an audio sample, and Causeur drives the behavior, so no human performer is needed.
What should I compare it with? Compare end-to-end latency, frames per second, identity stability and price per hour against the avatar services you already evaluate, using the same image and the same script.
The safety question
A system that builds a live, speaking character from one image and a short audio sample is powerful for tutoring and training. It is also a tool for impersonation. Before deploying, ask for consent checks on the face and the voice, watermarking or provenance labels on the output, and logging. Tell users when they are talking to a synthetic character.
The same concern drives the detection tools in the market. See our post on Pangram's new image and video detector for what detection can and cannot prove, and why provenance matters.
What is still unknown
- Pricing and rate limits. No public price list in the text we read.
- Quality evidence. The launch shows demos; we saw no blind comparison.
- Audio input length. 10 seconds versus 15 seconds across sources.
- Weights and license. Not open, as far as we can tell. Access is by invitation.
- Safeguards. We did not find a published policy on consent or watermarking for Atmee v2.
Bottom line
Atmee v2 is a plausible and interesting step toward video agents that talk live. A specialized, one-GPU design is the right engineering direction for cost. The 50x and 10 to 100x figures are claims from the company, compared against general video models, with no published methodology. Treat them as hypotheses to test when you get an invitation code.
Details reflect Atmanity's website and launch coverage as of November 4, 2026. Access, pricing and claims may change.