Search Results for "bit"

Z80-μLM is a retro-computing AI project that demonstrates a tiny language model (Z80-μLM) engineered to run on an 8-bit Z80 CPU by aggressively quantizing weights down to 2-bit precision. The repository provides a complete workflow where you train or fine-tune conversational models in Python, then export them into a format that can be executed on classic Z80 systems. A key deliverable is producing CP/M-compatible .COM binaries, enabling a genuinely vintage “chat with your computer” experience on real hardware or accurate emulators. ...

Watermark Anything (WAM) is an advanced deep learning framework for embedding and detecting localized watermarks in digital images. Developed by Facebook Research, it provides a robust, flexible system that allows users to insert one or multiple watermarks within selected image regions while maintaining visual quality and recoverability. Unlike traditional watermarking methods that rely on uniform embedding, WAM supports spatially localized watermarks, enabling targeted protection of...

...The flagship CogVLM-17B combines ~10B visual parameters with ~7B language parameters and supports 490×490 inputs; CogAgent-18B extends this to 1120×1120 and adds plan/next-action outputs plus grounded operation coordinates for GUI tasks. The repo provides multiple ways to run models (CLI, web demo, and OpenAI-Vision–style APIs), along with quantization options that reduce VRAM needs (e.g., 4-bit). It includes checkpoints for chat, base, and grounding variants, plus recipes for model-parallel inference and LoRA fine-tuning. The documentation covers task prompts for general dialogue, visual grounding (box→caption, caption→box, caption+boxes), and GUI agent workflows that produce structured actions with bounding boxes.

Qwen2.5-Omni is an end-to-end multimodal flagship model in the Qwen series by Alibaba Cloud, designed to process multiple modalities (text, images, audio, video) and generate responses both as text and natural speech in streaming real-time. It supports “Thinker-Talker” architecture, and introduces innovations for aligning modalities over time (for example synchronizing video/audio), robust speech generation, and low-VRAM/quantized versions to make usage more accessible. It holds...

...It provides scripts to download pretrained checkpoints of different model sizes (small, medium, large) trained on large-scale datasets and includes utilities for handling color quantization with a 9-bit palette. Researchers can use the code to sample new images, evaluate generative loss on datasets like ImageNet or CIFAR-10, and explore the impact of scaling on performance. While the repository is archived and provided as-is, it remains a valuable starting point for experimenting with autoregressive transformers applied directly to raw pixel data. ...

NVIDIA-Nemotron-3-Nano-30B-A3B-FP8 is a state-of-the-art large language model developed and released by NVIDIA as part of its Nemotron 3 family, optimized for high-efficiency inference and strong reasoning performance in open AI workloads. It is the post-trained and FP8-quantized variant of the Nemotron 3 Nano model, meaning its weights and activations are represented in 8-bit floating point (FP8) to dramatically reduce memory usage and computational cost while retaining high accuracy. The base Nano architecture uses a hybrid Mamba-Transformer Mixture-of-Experts (MoE) design, allowing the model to activate only a small fraction of its 31.6 billion parameters per token, which improves speed and efficiency without sacrificing quality on complex queries. ...

Command A+ 05-2026 W4A4 is a 4-bit quantized version of Cohere’s open-source Command A+ model, optimized for enterprise-grade agentic, multilingual, and reasoning-heavy workloads. It supports text and image inputs, generates text outputs, and uses a sparse Mixture-of-Experts Transformer architecture with 218B total parameters and 25B active parameters. The W4A4 release applies 4-bit weight and activation quantization mainly to MoE experts, preserving attention components at full precision to reduce quality loss while improving speed, latency, and hardware efficiency. ...

...The model supports a 256K-token context window, configurable thinking mode, native function calling, structured JSON output, and multilingual inference across 35+ languages. The NVFP4 quantization reduces weights and activations from 16-bit to 4-bit, lowering disk size and GPU memory needs for vLLM deployment.