AI-RADAR.IT

News & analysis on local LLMs, stack & on-prem hardware.

🗺️ Sitemap

Complete navigation guide to all content on AI-Radar.it

📱 Core Pages

🌟 Comprehensive Guides

🛠️ Tools & Features

🏢 LLM On-Premise

🏗️ Architecture

🏭 Industry Scenarios

📂 Article Categories

📡 Feeds & API

📰 Recent Articles

DeepSeek V4.1 Flash on Hugging Face: self-hosting back in the spotlight

NVIDIA's 4B BF16 and the VRAM bill in local inference

NVIDIA and the 4B BF16 Model: The Specific Weight of Local Inference

ORL pushes diffusion LLMs to 2,788 tokens/s on a single RTX 5090

Reflection's Beam: Token-Efficient Reasoning, 501B Memory Footprint

Hybrid Reasoning Agents: Hidden Thinking Cuts Token Use and Reshapes...

Tencent Hy-Memory-30B-A3B: Why 3B Active Parameters Matter More Than 30B Total

Qwen Flash Next: a solo Q2 quant challenges ISTA Lab

Intel Ships LLM-Scaler-vLLM PV 1.5: Docker Stack for LLMs on Arc GPUs

Open source sovereignty: the boring work of making it real

Bolt Graphics seeks $100 million to challenge Nvidia and AMD with the Zeus GPU

Meta’s Muse Spark Open: the real signal is in portable kernels, not 816B

PRISM: privileged information for multi-agent topology design

DeepUnfolding Agent: Adaptive Control Becomes an External Service

Saturated Benchmarks Still Hold Signal in Model Transcripts

DeepSeek-V4-Flash on 4x Radeon AI PRO R9700: the 79 tok/s configuration that...

Meta Opens Weights for Muse Spark Open: 816B Parameters and a 1M Token Context

Nemotron 3.5 Bolt: A 74B VLM Compressed to 24GB for On-Premise

RTX 5060 Ti hits $1,292 in Japan: GPU Black Friday is already a test

GNOME Mutter Improves Damage Tracking for Secondary GPU Copy Buffers

Microsoft's local 137B-parameter MoE model and the on-premise coding push

Mr.Chatterbox-2.7B-Chat: an LLM trained from scratch on Victorian texts

AMD W7900 and LFM2.5: A Concrete Signal for On-Premise LLM

Bengali shatters English LLM benchmark certainties

Following the Flow the Law Requires: Voice Agents and Regulated Debt Collection

Inverse Models Turn Answer-Only Data Into Reasoning Distillation

Lotus: forecasting as a pretraining principle for time-series LLMs

Artificiality's First Exam: An LLM-Referenced C-test for Supraintelligence

LiquidAI LFM2.5-24-A2B: Robust Performance on AMD W7900 for On-Premise LLMs

Ling-3.1-flash: 560B Parameters, 25B Active, and a 1M-Token Context Window

NVIDIA Nemotron Bolt 3.5 74B: 24GB MoE via 2-bit quantization

Skeleton Technologies surpasses $1B valuation with Siemens backing for US AI...

China's open-weight models trail US closed models by 7 months, and data...

Apple researchers turn idle Macs into a virtual GPU cloud with ‘Minions’

51 offline labs to learn AI from neurons to agents

Mesa brings YOLOX to NPUs: the open driver reshapes local inference

Mesa Expands NPU Support: YOLOX and PyTorch Join the Open Driver Stack

A jailbroken PS5 runs Qwen 27B at 20 tokens/s: a signal for self-hosting

LiquidAI LFM2.5-24B-A2B: 24B parameters, 2B active for on-device

Linux Kernel Developers Consider Dropping Assisted-By Tag for AI/LLM Patches

Google DeepMind’s resident philosopher and the empirical turn in LLM ethics

Reverse engineering CUDA with AI: the on-premise lock-in symptom

Reverse engineering CUDA: the ironic challenge from an AMD user

Intel's Crescent Island Becomes D-Series with Up to 480 GB LPDDR5X

From Crescent Island to D Series: Intel Targets Enterprise Inference

In 2030 a home GPU cluster will be seen like a gun for home defense

Intel D-Series D270 Targets Inference with FP8 and 1.5 TB/s

MiniCPM-V 4.7: the first open-source MoE with smooth on-device inference

Boole-1.7B: 1-bit quantization becomes a concrete on-premise option

ROCm 10 versus Rusticl on AMD Strix Halo: the open-source driver battle for...

View All Articles →