Gemma
PulseAugur coverage of Gemma — every cluster mentioning Gemma across labs, papers, and developer communities, ranked by signal.
14 天有情绪数据
-
Local AI tools boost LLM speeds with new prediction and decoding techniques
Recent updates in the local AI community are enhancing inference speeds and providing practical benchmarks for open-weight models. The llama.cpp project now supports Multi-Token Prediction (MTP), which has shown a 40% s…
-
Distributed output templates, not single positions, drive LLM in-context learning
Researchers have demonstrated that in-context learning in large language models is driven by distributed output templates rather than single-position activations. Through multi-position intervention, they achieved up to…
-
Users discuss chatting with Ollama or Gemma AI models
The user is asking if they can talk to Ollama or Gemma when feeling lonely, using hashtags related to AI.
-
Unsloth 推出用于本地 LLM 部署的 API 端点
Unsloth 发布了一个新的 API 推理端点，允许用户运行具有增强功能的本地大型语言模型。该端点同时支持 Anthropic 和 OpenAI 兼容的方言，从而能够与各种 AI 代理和聊天客户端无缝集成。此次更新还引入了 NVIDIA Nemotron 3 Nano Omni 和 Mistral 3.5 Medium 等新模型，并对 Unsloth Studio 进行了一些错误修复和改进。
-
Transformer models encode concepts in quiet spectral regions, syntax in high-variance ones
Researchers have identified a dual geometry within transformer representations, where concept directions anti-concentrate in the spectral tail while static unembedding-row contrasts concentrate in high-variance directio…
-
使用 LFM 2 和 Transformers.js，通过 WebGPU 在本地运行 LLM
Thomas Bley 发布了新的幻灯片，详细介绍了如何使用 LFM 2 在本地运行大型语言模型 (LLM)。该演示文稿还涵盖了将 Transformers.js 与 WebGPU 结合用于隐私过滤器、函数调用和嵌入，所有这些都在用户的浏览器中进行处理。
-
Developer builds complex AI system using no-code tools and existing models
A developer created a complex AI system without writing any code, leveraging existing Python and JavaScript modules, HTML overlays, and database tables. The system includes a desktop application with an installer, a Tel…
-
Curated learning path guides developers in building real-time voice AI agents
A new GitHub repository, "Voice-AI-for-Beginners," offers a structured learning path for developers to build real-time voice AI agents. The guide covers the entire process from initial speech-to-text calls to scaling pr…
-
AI safety research probes jailbreak success and emergent misalignment in LLMs
Two new research papers explore the underlying causes of AI safety failures in large language models. One paper introduces LOCA, a method to provide local, causal explanations for why specific jailbreak prompts succeed,…
-
IBM 发布 Granite 4.1 AI 模型系列，面向企业工作负载
IBM 推出了其 Granite 4.1 系列 AI 模型，这是其迄今为止规模最大的发布。这一新系列包括语言、视觉、语音、嵌入和 Guardian 模型，专为企业应用设计。这些模型旨在提高指令遵循、工具调用和转录准确性等方面的性能，并注重训练过程中的数据质量和分阶段优化。
-
New diagnostic tool probes LLM circuits for safety and behavior insights
A new research paper introduces "Perturbation Probing," a diagnostic method for understanding the internal workings of large language models. This technique uses two forward passes per prompt to identify and analyze "be…
-
AI model explores quaternion math for attention transformer architecture
A user explored the possibility of using quaternion algebra for attention transformers, conversing with a local Gemma 4:26b model. The model suggested it might be feasible and offer benefits, but warned that the inheren…
-
FlashNorm speeds up transformer inference by optimizing normalization layers
Researchers have developed FlashNorm, a technique to accelerate normalization layers in Transformer models. By reformulating RMSNorm and folding its weights into subsequent linear layers, FlashNorm enables parallel exec…
-
AI models struggle with emotion nuance, researchers explore new evaluation and generation methods
Researchers are exploring the nuances of emotion in AI, with several papers focusing on Large Language Models (LLMs) and speech processing. One study investigates how well small language models preserve emotions during …
-
LLMs show categorical perception and optimized data selection
Researchers have developed a new framework for optimizing data selection in large language models, adapting data weighting to specific tasks and models using efficient proxies. Another study investigates categorical per…
-
Google DeepMind unveils Decoupled DiLoCo for resilient AI model training
Google DeepMind has introduced Decoupled DiLoCo, a novel approach to training advanced AI models that enhances resilience and flexibility across data centers. This system can train models like Google's 12B Gemma model a…
-
CoreWeave enhances multi-cloud AI stack with Google Cloud interconnect and unified orchestration
CoreWeave has announced a suite of services aimed at simplifying multi-cloud AI infrastructure, including a direct interconnect with Google Cloud to reduce deployment times. The company also introduced SUNK Anywhere, a …
-
vLLM releases v0.19.1rc0 with Gemma 4 implementation updates
vLLM has released version 0.19.1rc0, which includes updates to its Gemma implementation. This release is part of ongoing development and feedback integration for the vLLM project.
-
Google DeepMind details 2025 AI breakthroughs with Gemini 3 and new models
Google DeepMind and Google Research have detailed significant AI advancements throughout 2025, highlighted by the release of their Gemini 3 and Gemini 3 Flash models. These models demonstrate state-of-the-art performanc…
-
Cactus launches open-source AI engine for mobile devices
Cactus has released an open-source AI engine designed for mobile devices and wearables, prioritizing low latency and reduced RAM usage. The engine supports multimodal capabilities, including speech, vision, and language…