Tag
AuroraAI-Research 发布了 AuroraIMG2-50M-Preview，一个基于 5040 万参数流匹配扩散 Transformer 的 256×256 文生图模型，搭配 1260 万参数文本编码器，全部从零训练仅需单卡 RTX PRO 6000 上 10 小时完成。
OpenBMB 发布 MiniCPM-OCR Jev（2B 参数），一款 OCR 质量评估模型，可对文本、公式和表格区域的识别结果打 0-1 分，在 JevOCRBench 上决策 AUROC 达 93.6，超过多个更大的模型，并在多种 OCR 模型输出上保持稳定泛化。
发布了一个仅 600 万参数的角色渲染模型，可从单张参考图生成约一分钟的视频，并与 Blender 渲染结果进行真值对比评测。
SteerTTS 是一个 23M 参数的可数值调控英文 TTS 模型，通过 z-score 控制语速、音高、响度等 12 项参数，可在 CPU 上实时运行，并附带 2825 个内置语音与实验性零样本克隆。
NVIDIA 发布 Nemotron Labs Dory1.0-4B-BF16，一款 4B 参数的循环式(latent-recurrent)大语言模型，采用 normalized-GPT 架构与滑窗/全注意力混合层，支持弹性推理计算与最后一次循环 KV 缓存，并原生支持 NVFP4 预训练，采用 OpenMDW-1.1 开放许可。
AuroraIMG-6M 是一个仅 6.82M 参数的文本到图像模型（含文本编码器、去噪器和 VAE），全部从零在 CC12M 上训练，可生成 64×64 图像，FID 33、CLIP 24.8，甚至能在 CPU 上运行。
The author describes a second-opinion checker pattern for finance approval agents and evaluates webAI's new 3.66B logic-only model TwIL-LM3 Pro, which matches Qwen3-8B's logic scores at less than half the size and runs locally for data privacy.
Anthropic 发布 Claude Haiku 5.5，宣称是其最便宜、最快且最强的小模型，相比 Haiku 4.5 运行成本降低约 75%，并支持 100k token 上下文窗口。
Anthropic 发布 Claude Haiku 5.5，这是其最快、能力最强的小型模型，专为摘要、分类、编码子代理、客户支持和浏览器使用等高容量、成本敏感的任务而设计。
Google DeepMind unveiled a 740M-parameter multimodal model that is competitive across benchmarks, outperforming some specialist models more than twice its size, enabling developers to add multimodal search features like finding video moments via voice memo and pair it with Gemma 4 for private on-device use.
DecisionTune 1.0 is a 395M-parameter encoder (ModernBERT-large plus a scoring head) that scores options offline for small agent decisions, running in ~10 ms per decision on MLX with Apache-2.0 weights and backends for PyTorch, MLX, and ONNX.
A user tested a 6B-parameter small model on a full poster and praised its ability to break down every layer — title, background, and main elements — while the quoted post shares a cloud deployment tutorial for running open-source image/video models at roughly 1/8 of the official API price.
Ant Group and Baidu open-sourced Ming-Image-0.1-Design, a compact 6B-parameter image generation model that topped Artificial Analysis's open-source UI/UX design leaderboard, along with a companion tool Layer that decomposes generated images into editable transparent layers.
Nace AI released Drex 1.5, a decision-focused model under 10B parameters with a 128k context window, ranked #1 on the Decision Index 0.2.1 with a score of 58.28, priced at $0.04 per 1M input tokens.
webAI introduces TwIL-LM3-Pro, a 3.6B-parameter open-source reasoning model that runs locally on everyday computers with quantized builds, claiming top small-model scores on formal logic, SVAMP, MuSR, and BIG-Bench Hard. They also tease Meridian, an upcoming family of frontier-class on-device models.
A developer built a fully offline AI assistant on a Mac mini using a 3B parameter model that can diagnose and fix code errors via voice input with zero API costs and no internet, sharing the full build and repo.
This article provides an update on the Engram model, detailing its 2.6b parameter architecture with a large Engram table and initial training progress at 100m tokens, showing improved completions with context-aware data offloading.
The paper details ufakzeka-1, a 151M-parameter Turkish language model built from scratch with a total cost of about $286, describing the training pipeline, evaluation methods, and key findings on small-model training limitations.
AgentCloak is a free browser extension that uses a small AI model to locally detect and replace personal details in AI prompts, ensuring privacy across services like ChatGPT and Claude.
Cua has open-sourced CUA-S1-FORMS, a tiny 2.8MB AI model specialized for form-filling tasks, achieving 99.7% accuracy and enabling local deployment.