media Hugging Face Forums · 2d ago · open_models

NanoAvatar enables local talking avatars on Android via Qualcomm QNN

from English

NanoAvatar is an open-source project that runs lifelike talking avatars locally on older Android phones without requiring a cloud GPU. The release includes public code, model weights, and Android APKs that bundle the model and an avatar for offline use with a user's own voice.

The Android runtime utilizes Qualcomm QNN for inference.
Full and Lite versions share one codebase but bundle different model and avatar resources.
On a Z60 Ultra with Snapdragon 8 Gen 3, the Lite version reached 37 FPS with an 112 ms model first frame.
Released APKs generate video at a fixed 25 FPS for Full and 12.5 FPS for Lite.

The author welcomes feedback from developers working with QNN or integrating talking avatars into their applications.

Importance 1/3 Trust 1/3 Hugging Face Forums Local inference Video generation

Read original

github llama.cpp · 3h ago · 4 views

llama.cpp b10982 adds Vulkan sparse Flash Attention support for DSV4/GLM

The llama.cpp project has released build b10982, which introduces support for sparse Flash Attention via the Vulkan backend. This update specifically targets DSV4 and GLM models, allowing them to leverage sparse attention mechanisms on compatible GPU hardware.

github llama.cpp · 4h ago · 2 views

llama.cpp b10981 fixes stateful decode for Gemma-4 and optimizes GPU MoE inference

The llama.cpp b10981 release addresses critical correctness issues in the OpenVINO backend's stateful decoding path, specifically fixing support for models like Gemma-4 that use per-layer head sizes. It also introduces optimizations for GPU Mixture of Experts (MoE) inference and improves performance through KV state relayouts.

lab NVIDIA Research · 16h ago · 3 views

FastGen-PDD introduces parallel decoding distillation for fast video and image generation

Researchers introduce Parallel Decoding Distillation (PDD), a simplified trajectory-based method to accelerate inference in diffusion and flow matching models. Unlike current approaches that rely on hard-to-optimize variational score distillation and adversarial losses, PDD predicts multiple denoising steps per network evaluation.

github llama.cpp · 21h ago · 8 views

llama.cpp 0.4.1 adds Maple 20B-A1B, Tencent Hy 4, Spark2.5 support

The llama.cpp project released version 0.4.1, introducing support for the Maple 20B-A1B ternary Mixture of Experts architecture on CPU, as well as preview support for Tencent Hy 4 and Spark2.5 models.

media Hugging Face Forums · 1d ago · 2 views

Discussion on practical limits of running 2B–7B models locally

The article poses a question regarding the practical utility of smaller open-weight models (2B–7B) for local deployment, asking whether they are becoming sufficient for personal workflows compared to larger models.