Articles

DeepSeek released V3-0322, an open-source MoE model with 671B total / 37B active parameters that matches GPT-4.5 on key benchmarks while remaining fully self-hostable under MIT license.

Research shows LLMs can deanonymize pseudonymous users across platforms with 90% precision and 68% recall, threatening the fundamental assumption that pseudonymity provides adequate online privacy.

A new research paper documents 'Cross-Model Void Convergence' — a phenomenon where GPT-5.2, Claude Opus 4.6, and other frontier models independently converge on identical patterns of structured non-re

Tinybox is a portable, fully offline AI device capable of running large language models with up to 120 billion parameters. It represents a significant step toward private, edge AI computing that doesn

A new HomeSec-Bench benchmark shows Qwen3.5-9B running locally on a MacBook Pro M5 achieves 93.8% accuracy on security tasks, just 4 points behind GPT-5.4 cloud — with zero API costs and full privacy.

Haskell expert Gabriel Gonzalez dismantles two core misconceptions behind agentic coding hype: that specs are simpler than code, and that writing specs is more thoughtful than coding. Using OpenAI's S

小米在 3 月 19 日发布 MiMo-V2-Pro / Omni / TTS 系列模型。此前一周，一个未署名的 1T 参数 Hunter Alpha 模型在 AI 社区引发轰动，被广泛猜测为 DeepSeek V4，最终被小米认领。同时发布基于 Omni 模型的「龙虾」QClaw AI Agent。

ICML 2026 detected that 506 reviewers assigned to 'no LLM' policy violated their own commitment, generating 795 AI-written reviews. The conference desk-rejected 497 corresponding author submissions an

A researcher topped the HuggingFace Open LLM Leaderboard not by training, merging weights, or fine-tuning — but by duplicating 7 middle layers in a 72B model. The discovery reveals that LLM middle lay

Open-source 14-stage compression pipeline achieves 54% average token reduction across code, JSON, logs, and agent conversations with zero LLM inference cost. Outperforms LLMLingua-2 by up to 88% at ag

Unsloth launches Studio — an open-source, no-code interface for training, running and exporting 500+ open models locally with 2x faster training and 70% less VRAM, powered by NVIDIA DataDesigner for a

A mysterious 1T-parameter model called Hunter Alpha appeared on OpenRouter, sparking speculation that DeepSeek is quietly testing its V4 model.

OnPrem.LLM AgentExecutor: 2-line autonomous AI agent with 9 built-in tools (file/shell/web), sandboxed execution, cloud+local model support.

OpenAI releases GPT-5.4 mini (2x faster than 5 mini, 400K context, /usr/bin/zsh.75/1M input) and nano for high-volume workloads — designed for subagents, coding, and computer use.

Research proposes distributed systems theory as a principled framework for LLM agent teams — drawing parallels between multi-agent AI and classical distributed computing.

Hello everyone, I've been working on mlx-tune , an open-source library for fine-tuning LLMs natively on Apple Silicon using MLX. I built this because I use Unsloth daily on cloud GPUs, but wanted to p