Apple's AI researchers publish papers on using human gaussian splats to generate animated 3D avatars and deploying LLMs on devices with limited memory
Apple, a company practically synonymous with technological innovation, has once again positioned itself at the forefront of the AI revolution.
VentureBeatMichael Nuñez
Context & Ripple Effects
Apple had already begun opening its AI work to public scrutiny with its first public AI research paper, making these publications part of a longer shift from a historically closed research posture.
The two papers also foreshadow later coverage of Keyframer’s prompt-based image animation and a strategy combining local and cloud language-model processing in Project Greymatter. Together, they connect visual generation research with the practical constraints of device-resident AI.
First-order effects
Apple publicly documents work in two distinct technical areas: animated 3D human representations and language-model operation under tight memory limits, giving external researchers concrete methods to assess and build on.
The reporting describes research rather than a product launch, so there is no immediate announced change for Apple customers or developers.
Second-order effects
The limited-memory LLM work strengthens the technical case for product teams to prioritize efficient local inference alongside cloud processing, consistent with the later local-and-cloud AI architecture coverage.
The avatar work and subsequent natural-language animation research create a more connected research path for generative visual tools, though the corpus does not establish a commercial product.
Third-order effects
If Apple continues to pair public research with local-and-cloud AI design, competition may increasingly hinge on optimizing models for hardware constraints rather than relying solely on larger cloud models.
Public publication can also serve as a talent and ecosystem signal: firms developing proprietary AI hardware have reason to demonstrate technical leadership while retaining control over product timing.
The trend: This is one data point in the shift toward hybrid AI systems, where generative capabilities are designed around both device limits and cloud-scale processing.
Introducing 🫂HUGS: Human Gaussian Splats - capable of creating animatable (3DGS) avatars from a casual video (50-100 frames) in ~30 mins. Our avatars can easily be embedded into other (NeRF) scenes. (1/4) Project: https://machinelearning.apple.com/ ... arXiv: https://arxiv.org/..…
And so it begins: @Apple announces LLM in a flash: Efficient Large Language Model Inference with Limited Memory. Brilliant move! paper page: https://huggingface.co/... LLMs are a new kernel and should be a low-level utility embedded in every device that can be updated over...
My bet: we are heading to a XLA-like compute graph optimizer, which will be (1) able to deal with CPU-GPU, (2) will be aware of activation statistics, (3) will do approximate compute, including quantization.
@gabrielnocode @Apple Apple introducing LLM in a flash is like giving your phone a PhD. Can't wait for my iPhone to explain quantum physics to me... or at least help me order pizza more efficiently.
problems I see with this paper: - they only consider under trained models: falcon7b and opt7b, so they can get 90% sparsity (that won't work with mistral7b) - they credit the speed up for sparsity to their method and compare inference time to a dense model (why?!) (1/2)
Further evidence that Apple will emerge as a big player in AI. Maybe not in the flashy big models, but in how the majority of people will interact with GenAI - on device.
The research team at Apple revealed Human Gaussian Splats, or ‘Hugs’. The AI avatar research can now take short 2D video clips and take a 3D avatar from it. In the future, when people take a recording of you, they can put you anywhere. https://x.com/...
Apple AI avatar research that takes your short 2D video clip and pulls apart an editable 3D avatar and environment. The modelling of hair, clothing etc is immaculate and looks far more realistic than rigging a 3D photo scan.
Apple announces LLM in a flash: Efficient Large Language Model Inference with Limited Memory paper page: https://huggingface.co/... Large language models (LLMs) are central to modern natural language processing, delivering exceptional performance in various tasks. However, their.…