AI Briefing#model-architecture
The latest AI and developer news about #model-architecture, with the original source and a short summary.
The latest AI and developer news about #model-architecture, with the original source and a short summary.
The model features a Sparse MoE architecture with 14B active parameters per token and outperforms DeepSeek-V4-Flash-Base on MMLU and coding benchmarks.
ViKernel v0.4 transplanted CPU, NPU, GPU, RAM, and ROM circuits into the Meta Muse Glimmer 30B model weights as a sidecar layer, effectively blocking agent security threats.
The method compresses the first loop visit and reallocates saved computation to increase hidden size across 1.3B-7B parameter models.
Cross-layer shared Meta Experts reduce parameter redundancy and improve downstream task accuracy compared to hierarchical Local Experts.
Linum AI proposes Pyramid-JiT, a VAE-free pixel-space text-to-image model that achieves an FD-DINOv2 score of 27.9, approaching the real-image floor of 25.4, while requiring 11.3x fewer training samples than Linum v2.
CLM-8B achieves SOTA on agent coding benchmarks by separately encoding and caching states and actions.
Kuleshov's research team summarized diffusion language model (dLLM) technology that overcomes the speed limits of AR models through parallel generation, along with the current status of commercialization efforts such as Mercury 2.
New research demonstrates that combining diverse mechanisms (especially including SSMs) has a greater impact on performance than the placement order of attention mechanisms in Transformer stacks.
Researchers from the National University of Singapore and others released OpenWAM, a modular stack for comparing design variables of robot World-Action Models.
A model fine-tuned on image deraining tasks using the depth estimation backbone and neck weights of YOLO26 has been released.
The GLM-5.3 model has been visualized on HF Viewer, confirming that while its architecture remains identical to GLM-5.2, performance has improved through training enhancements.
Qwen4Exp maintains inference speed by offloading 25% of weights to SSD using N-grams.
Analyzed the vision tower structure and benchmark performance characteristics of the Qwen3.8-27B model.
AWS optimizes AI models and kernels alongside chip design to strengthen Trainium's competitiveness.
AFM3 20B is a MoE model that selects only about 20% of MLP layers to activate per prompt.
Technical specifications have been released for Neutrino-1 8B, a dense decoder-only transformer model built on Qwen3-8B.
Liquid AI has unveiled a technique that expands the vocabulary of the LFM2.5-8B-A1B model from 65K to 128K without retraining from scratch.
It solves the single-stream bottleneck of LLMs by separating thoughts, inputs, and outputs.
A Residual Coupling technique is proposed that connects fixed language models via linear bridges to boost performance without modifying weights.
Orthrus, based on Qwen3, achieved up to 7.8x speedup through lossless parallel generation.
Thinking Machines released a research preview of Interaction Models for real-time multimodal interaction.
Interfaze has unveiled a new model architecture and API for OCR, STT, and structured output.
SATFormer is proposed, improving efficiency by selectively reusing early-layer representations on a per-token and per-head basis.
SenseTime has open-sourced the SenseNova U1 multimodal model based on NEO-Unify.