AI Article Tags

Browse articles by professionally maintained tags. Every tag opens a focused collection with its own editorial introduction and closely related published articles.

Xiaomi MiMo 309B total parameters with 15B active: Xiaomi MiMo uses MoE to push cost down while stretching context to 1M. A 7B reasoning series started the line, V2-Flash, V2-Pro, Omni and TTS extend it into a full agent stack, and V2.5 ships open under MIT. Open WebUI A self-hosted AI front door rather than another chat box. Ollama, vLLM and OpenAI-compatible services land in one layer with accounts, permissions and knowledge bases; Docker is the easy starting point, and the branding terms deserve a look. OpenRouter One key reaches hundreds of models, billed by usage, with automatic fallback when the primary provider fails. Whether it plugs into your own tools depends on the API being OpenAI-compatible and the base_url being right. vLLM Throughput, memory and concurrency are the three things vLLM cares about. There is no chat interface; it turns open models into an API that holds up under production traffic, while tool calling, MoE backends and long-context caching keep improving. Agent Skills One SKILL.md, a trigger description and a few optional scripts turn a working method into a shareable capability. Skill libraries now circulate between Codex, Hermes and OpenHands, so check the source, the permissions and what the scripts do before installing. Kiro Let the AI write the requirements first and the code second: Kiro reverses the order. Specs, steering and agent hooks are its core trio, and powers bundle MCP servers with rules into packages loaded only when a task needs them. Meituan LongCat 560B total parameters with only 27B active, plus 100,000 free tokens a day: LongCat pushes the barrier down to cost. ScMoE and asynchronous RL are its two main threads, and the models now reach from text to image, speech and video. Mistral AI Open the weights, keep compute and delivery in hand: Mistral picks a different route. The Mistral 3 family ships under Apache 2.0, Ministral covers 3B to 14B for the edge, and Studio turns prompts into rollback-ready assets. Baidu Wenxin Wenxin became known as a Chinese chat assistant and now splits into base models, a coding assistant, image generation and document parsing. ERNIE 5.0 moves to native omni-modal modelling while ERNIE-4.5-VL and PaddleOCR-VL open their weights under Apache-2.0. Tencent Hunyuan One lab put language, image, video, OCR, motion and translation models onto open platforms within a year. Parameter counts run from 1B to 770B, most weights ship as Apache-2.0, and the modality spread is wider than most vendors manage. Quadruped Robots Steps, gravel and refinery stairways that defeat wheels are exactly where legs win. This sets out why consumer models compete on price and openness while industrial ones compete on ingress rating and inspection payloads, and why the use case comes first. Computer-Use Agents Legacy systems with no API finally get an automation path. This covers how computer-use agents read the screen, locate controls and click, and why permission prompts and sandboxes have to ship before the capability does. AI Voice Agent Only at roughly a hundred milliseconds of latency does speech start to feel like conversation. This covers which conversions end-to-end speech-to-speech skips compared with the ASR to TTS chain, and how turn-taking plus tool use move voice from talking to getting things done. AI Guardrails Asking a model to police itself is not a safety boundary. This breaks down the layers a guardrail stack needs, where prompt injection gets in, and why permissions and approvals must live outside the model in deterministic code. LLM Inference Optimization Change how requests are batched and the same GPU can serve several times the throughput. This explains what continuous batching, speculative decoding, KV cache and context caching each save, and how far kernel libraries and sparse architectures can push it. AI Workplace Collaboration AI inside office suites is leaving the chat box: Copilot picks up tasks in Word, Teams and Outlook, Dingtalk pushes work forward with Agent OS, while Notion AI and Zapier cover knowledge and cross-app actions. AI Frontend Development Turning a screenshot or a single prompt into a running page is the most visible shift in frontend tooling: where Redesign Agent and LlamaCoder stop, what backend Chef ships, and what is left before shipping. AI Social Media Many accounts, a publishing calendar and honest reporting are the three real hurdles: which scale each of Buffer, Hootsuite, Later and Metricool fits, and who handles topic research and layout on Chinese platforms. AI Product Management Along the path from insight to prioritisation to prototype validation: which drafting work AI has taken over, and which trade-offs stay with the product manager. AI Code Review What automated review actually produces: how CodeRabbit and Sourcery build line-by-line suggestions, which context MCP brings in, and why Copilot needs a fresh request after new commits. AI Observability What happens after an AI app ships: how tracing, cost attribution and latency metrics locate failures, and what a gateway product adds on top. AI Sales Around three moments - finding leads, writing outreach and advancing deals - a comparison of Clay, Apollo and Agentforce, and what happens when the fit is wrong. AI Content Detection The three ways to spot AI-generated content: probability scores from detectors, platform likeness detection, and watermark provenance, plus what a false call costs. AI for Manufacturing How AI is actually used on the factory floor: engineering simulation and design optimization, industrial agents and inspection, plus the constraint of keeping data on site.