Tag
Recommends the locally deployable face swap LoRA "BFS-Best-Face-Swap," supporting face, head, and body replacement while preserving the original hair, skin texture, and lighting. Provides ComfyUI workflows, Draw Things presets, and Python Diffusers deployment options.
I recommend Manning’s new book, *Domain-Specific Small Language Models: Efficient AI for local deployment* (by Guglielmo Iozzia, published in May 2026), which explains how to fine-tune small models on domain-specific data, quantize them, and deploy them on local hardware to achieve high-quality performance on specialized tasks. It also covers RAG and agent systems that rely solely on small models.
Open-source tool Toonflow automatically generates AI short dramas in under two hours for around 120 yuan, covering the full pipeline from screenwriting and storyboarding to character design and final video. It uses a three-layer agent architecture with local deployment support, and has already reached 16.6k stars on GitHub.
JetBrains Mellum 2.1 12B A2.5B MoE 被评价为一款表现优秀的编码模型，能在 M4 Max 上以 130k 全上下文跑出 60 TPS，在 9070XT 上达到 140 TPS，且可在 8GB/12GB 显卡上流畅运行。
Manning released "Domain-Specific Small Language Models: Efficient AI for local deployment" by Guglielmo Iozzia, covering how to build, fine-tune, evaluate, and deploy compact AI models tailored to specific industries and workflows for lower latency, cost, and better data control.
作者系统整理了本地部署模型的加速方案，包括在 macOS 上结合 Turbo、MLX、VAE Q8 和前缀 KV cache 让 Qwen-Image 2.1 提速近 5 倍，以及用 Uzu 的投机解码把 Qwen3.5-9B 从 22 tok/s 提升到 92 tok/s、用 TensorRT 让 Qwen 在 RTX 5090 上比 llama.cpp 快 3 倍等案例。
An introduction to AtomicChat's "decensored" text encoder for Qwen-Image-2.1-Turbo, released on HuggingFace in GGUF quantized form, plus a complete tutorial for running it locally with ComfyUI on 8GB VRAM (RTX 5060). Includes the full file download checklist, node configurations, 5 real-world gotchas with their fixes, and prompt examples.
An AI content creator reflects on reaching 40k followers after a year, discussing the trade-offs of commercial sponsorships, investing earnings into local GPU setups (4090, dual V100s, a planned 4-card rig) for 'token freedom', and why future value lies in what AI cannot do.
一款基于27B Qwen修改的新版本模型发布，通过EfficientThink优化推理过程，减少多达53%的推理token，LCB代码基准从83提升至90，性能据称可比肩Opus5.5和GPT-6，支持Apache 2.0许可和本地部署。
本文提出一个基于大语言模型的框架，利用微调的开源 LLM 自动化生成交通管理计划（TMP），并在威斯康星州交通部的 WisTMP 系统中进行案例研究，发现微调显著提升性能，但模型仍难以生成项目特定理由和准确的成本估算。
Appreciation post for tcclaviger adding expert RAM offloading support to vLLM, enabling running large AI models like DeepSeek-V4-Flash-Vision-Exp on local setups with multiple GPUs.
A new Cyber-OSINT AI model is released, featuring a MoE architecture trained on OSINT instructions for local execution and cyber threat intelligence tasks like attribution and geolocation.
This post introduces Cyber-Ornith-1.5-9B-OBLITERATED, a 9-billion parameter AI model for cybersecurity that runs locally on laptops, trained on cybersecurity trajectories with zero refusal and strong performance in triage and tool-calling.
The author trained a small language model to replace Gemini Flash for a summarization task, achieving 97% accuracy with 0.06s latency, suitable for deployment in an internal app.
A developer shares their achievement of reducing object detection latency to below 0.35 seconds on local hardware, highlighting progress in AI performance optimization.
Discussing the image generation speed of Qwen-Image-2.1 local deployment, congratulating its release, and highlighting its breakthroughs in small parameter counts and high efficiency, positioning it as a potential leader in domestic AI image generation.
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.
Introducing Ternary Bonsai 2 27B, a highly compressed AI model that retains 98.2% of performance while being 9x smaller in footprint, enabling efficient local deployment for tasks like reasoning, coding, and multimodal processing.
llama.cpp releases version b11003, a C/C++ tool for efficient large language model inference with minimal setup on a wide range of hardware.
This paper presents BudgetBench, a protocol and harness for evaluating memory strategies in local large language model agents by varying per-call input-token budgets, offering a reusable measurement surface for community benchmarking.