on-device

NVIDIA released Nemotron 3.5 Bolt E4B 4GB, a post-trained Gemma 4 E4B model with multi-turn agentic and conversational capabilities optimized for running on RTX, Jetson, and DGX Spark devices, claiming ~40% higher quality and ~80% fewer output tokens than peers.

Arc-Transcriber is a ~70MB streaming speech model that performs ASR and NER/entity-linking in a single step, outputting transcripts, linked names, numbers, speaker roles, and audio events every 100ms via trained heads on a shared encoder, eliminating the usual multi-model pipeline. Two variants (Arc mini and Arc-v1) are released with weights on Hugging Face and code on GitHub.

Xenova 宣布突破：首个 100% 本地在浏览器中运行的真正全模态模型 LFM3-Omni-2B，借助自定义 WebGPU kernels 实现约 280 tok/s，支持视觉、音频和文本输入并生成交错的文本与音频输出。

OpenBMB released MiniCPM-OCR 1.0 (1B), a lightweight OCR model that runs locally on iPhone/iPad, plus MiniCPM-OCR-Jev (2B), a verifier that scores OCR output correctness against the source image; both ship with open weights and support llama.cpp, vLLM, SGLang, and Ollama.

Liquid AI 发布 LFM3-Omni-2B 的 GGUF 版本，支持在 CPU/GPU 上本地推理。该模型可理解文本、图像、视频和音频，并以文本和语音输出，专为设备端使用打造。

Liquid AI 工程师展示 LFM3-Audio 模型在本地设备（一盏灯）上运行并进行语音对话，网友戏称"能否在灯上跑模型"正成为新的基准。

Liquid AI 发布第三代 Liquid Foundation Models（LFM3），首批包含六个 2B 级模型，覆盖文本、视觉、音频、全模态和决策任务，采用开放权重并共享主干架构，可在手机、笔记本和边缘设备上运行，权重已上线 Hugging Face。

Liquid AI 推出新一代 Liquid Foundation Models 系列 LFM3，采用共享骨干的多模态（文本、视觉、音频、全模态、决策）混合架构，首批 2B 参数级模型开放权重，可本地微调并跨设备运行。

NotiKeep 是一款 iPhone 通知历史记录应用，利用 iOS 27 Shortcuts 自动化在本地保存通知，支持查看已删除或已编辑的消息、按应用与会话搜索，以及关键词提醒，全程本地处理、无需账号。

Block released Dessa, a full-duplex speech-to-speech voice model trained entirely on human speech that runs locally on consumer GPUs and produces expressive, human-like conversation with built-in emotion.

Liquid AI released LFM3-1.7B, a 1.73B-parameter tri-hybrid text model designed for local agents and edge deployment with a 256K-token context window, fitting in about 1 GB of memory at 4-bit precision and supporting single-GPU fine-tuning for domain specialization.

A free in-browser TTS tool that runs entirely on-device using a 2.7 MB model on a custom 136 kB WebAssembly engine, making Kokoro-distilled Paradee 12.9x faster while maintaining quality. The author breaks down each optimization step and lets users benchmark it live in their browser.

WhissleAI released SteerTTS-Edge, a steerable small English TTS model compressed to 11.84 MB via int4 quantization-aware training of the acoustic model and a distilled Vocos vocoder, with unchanged intelligibility (WER 0.96%) and only a ~0.1 UTMOS drop while preserving all 12 steering controls.

作者在 MacBook Pro (M5 Pro, 64GB) 上用 MPS 后端运行 0.9B 的 MOSS-Transcribe-Diarize 模型，对 AMI 测试集一段 17.5 分钟、4 人会议做单次转写+说话人分离，仅用 12 分钟（约 0.68x 实时）完成，4 位说话人全部识别，说话人混淆率仅 0.08%，DER 为 15.4%，并能正确处理重叠对话。

SparkLLM 釋出 Spark-X2.6-4B 端側模型，具備 4B 參數、Apache 2.0 授權、最高 1M context 與 3 種思考模式，並支援 GGUF 本地部署。

Liquid AI 发布 LFM3-Omni-2B，其 LFM3 家族首款全模态模型，可理解文本、图像、视频和音频，并以文本和语音作答，专为手机、笔记本等边缘设备设计，2.2B 参数即可替代独立的 ASR、VLM、LLM 和 TTS 模型。

Alkove is an open-source (Apache-2.0) on-device voice front-end made of five tiny models (7.5 MB total) — a typed wake-word system, voiceprint-based speaker filtering, and personal VAD — running on a custom Rust inference engine with npm, pip, Android, LiveKit, and Pipecat packages, aimed at reducing false wake-ups and cross-talk before STT in local voice assistants.

Liquid AI 发布 LFM3-VL-2B，一款面向端侧视觉理解的 1.92B 视觉语言模型，支持文本、图像和视频输入，并引入视频理解、逐步思考以及空间与具身推理能力。

Spark-X2.6-4B is a compact 4B on-device language model released in GGUF format, featuring improved post-training for coding, agentic, long-context, and multilingual tasks along with a controllable thinking budget. It claims state-of-the-art performance among similarly sized open-source models, matching or surpassing ~9B parameter models.

Muse Glimmer 是一款 30B 参数的开源多模态智能体模型，通过 4-bit 量化可在消费级硬件上本地运行多步推理与工具调用，并在同尺寸级别 agentic 基准上超越 Gemma4-31B 和 Qwen3.6-27B，现以 Apache 2.0 协议发布。