Tag
Google 推出 EmbeddingGemma 2，一个统一文本、图像、音频与视频表示的开源多模态嵌入模型，采用简化逐层嵌入与多阶段训练课程，在各模态基准上建立新的帕累托前沿，性能超越大 2 至 12 倍的模型。
Kyutai Labs introduces REKI, a 2.4B vision-language-action model pretrained exclusively on public data that sets a new Pareto frontier between performance and model size, using a simple design with register-token memory in the action expert. Full code, data curation pipelines, and all pretrained/finetuned weights are released.
MIT 6.5940: TinyML and Efficient AI Computing, taught by MIT Associate Professor Song Han, has open-sourced all its slides and videos — covering model compression, TinyML deployment, LLM and diffusion acceleration, and efficient training, with five hands-on labs.
LightMIS presents a scalable family of ultra-lightweight convolutional networks for 2D binary medical image segmentation, achieving high accuracy with reduced parameters and GFLOPs, suitable for on-device execution.
QVAC Genesis III is an open-source synthetic STEM corpus designed to enhance language model pre-training efficiency, demonstrating significant benchmark improvements over prior datasets.
Empero AI releases their first MoE model, Qwen3.8-35B-A3B-Distill, which distills reasoning from Qwen3.8 into a smaller, efficient architecture, showing improved benchmarks on ARC-Challenge.
LeVJEPA introduces a more efficient self-supervised video pretraining approach by eliminating the need for target networks and predictors, using a single shared encoder with SIGReg regularizer, and achieving competitive performance with lower compute costs.
The paper introduces BF1, a causal dyadic sparse-attention retrofit designed to improve the efficiency of transformers for long-context processing.
GRNEdit is a lightweight two-stage framework for efficient general video editing that uses binary semantic decisions and source evidence to achieve high performance with fewer parameters than larger models.
DistilVDR is a compact 524M visual document retriever distilled from an 8B teacher via cosine alignment, achieving near-teacher accuracy on ViDoRe with 15.6x smaller indexes and faster indexing.
This paper from the Kathleen series shows that an attention-free, byte-level model with ~0.5M parameters can beat a parameter-matched transformer on WikiText-103 language modeling and generation, introduces a non-parametric 'Form Distance' metric for evaluating text realism, and demonstrates that retrieval-augmented decoding from the model's own training corpus improves generation quality.
AntLingAGI announces Ling 3.0 flash, a native hybrid-linear reasoning model with 124B total parameters and 5.1B active, MIT-licensed, claiming to match a 1T flagship with much less compute and faster response times.
Maple-Preview is a ternary 20B MoE model that runs at 120 tokens per second on an iPhone, showcasing efficient on-device inference.
A user asks about Mach-1 Additive, a model reportedly achieving 95% of Qwen 3.6 35B's performance while being 10x smaller, questioning why it isn't more widely discussed.
It is claimed that DeepSeek V4 Flash has only 284 billion parameters and only 13 billion active parameters, posing an efficiency shock to model manufacturers in China and the US.
Thinking Machines released Inkling-Small, an efficient open-weights Mixture-of-Experts model with 276B total and 12B active parameters, achieving comparable performance to its larger sibling Inkling at a quarter of the size. It features native reasoning over audio and images, variable thinking effort, and a 1M-token context window.
Liquid AI releases LFM2.5-Encoders (230M and 350M), efficient encoder models optimized for long-context inference on CPU, matching or beating larger encoders on benchmarks with 3.7x speedup over ModernBERT-base.
Google's Gemma 4 26B A4B model can run on iPhone 17 Pro using model paging, enabling powerful on-device AI.
Four Over Six (4/6) introduces adaptive block scaling for NVFP4 quantization, reducing quantization error with minimal overhead, improving both training and post-training quantization for large language models.
Ant Group released LingBot-Vision, a family of DINO-style vision backbones in 4 sizes; the 0.3B ViT-L matches DINOv3-7B on NYUv2 depth with ~23x fewer parameters, showcasing significant efficiency gains.