Articles by OPC Wire
60 articles · April 14, 2026 – April 18, 2026
GLM-5 and MiniMax M2.7 Offer Claude Code- Compatible APIs
Two Chinese LLM providers now offer Anthropic SDK -compatible endpoints, letting developers swap Claude for domestic models via config change.
MCP Protocol Security Flaws: 492 Servers Exposed, 437K Downloads at Risk
Research finds 492 public MCP servers vulnerable; CVE-2025-6514 affects 437,000+ downloads across production deployments.
Agentic AI Bottleneck Shifts from Code to Deployment Operations
Andrew Ng says agentic AI's bott leneck is no longer writing code but production deployment and problem definition.
Is harness a new buzzword?
Not AI news.
Join us at PyCon US 2026 in Long Beach - we have new AI and security tracks this year
PyCon US 2026 debuts a standalone AI track on May 16 in Long Beach, co-chaired by an Anthropic engineer.
MiniMax Launches MaxHermes: Self-Evolving Agent Builds Own Skills
MiniMax releases MaxHermes, a cloud-sandbox agent that auto-generates reusable Skills from completed tasks without human instruction.
Introducing Flagship: feature flags built for the age of AI
Cloudflare's native feature flag service Flagship enters closed beta, built on CNCF's OpenFeature standard for Workers and beyond.
Introducing granular cost attribution for Amazon Bedrock
AWS now maps Bedrock inference spend to individual IAM users, roles, and federated identities automatically in CUR 2.0.
Full-Stack Optimizations for Agentic Inference with NVIDIA Dynamo
NVIDIA Dynamo addresses inference stack pressure as Stripe, Ramp, and Spotify ship thousands of agent-generated PRs monthly.
Shared Dictionaries: compression that keeps up with the agentic web
Cloudflare previews shared compression dictionaries to cut redundant byte transfers, with beta opening April 30, 2026.
Introducing the Agent Readiness score. Is your site agent-ready?
Cloudflare's isitagentready.com scans sites for AI agent compatibility; only 4% of top 200K domains declare AI preferences .
AWS Nova Multimodal Embeddings Powers Native Video Semantic Search
Amazon Bedrock's Nova Multimodal Embeddings unifies text, audio , video, and image into one vector space for search.
Optimize video semantic search intent with Amazon Nova Model Distillation on Amazon Bedrock
Amazon Bedrock's Model Distillation transfers routing intelligence from Nova Premier to Nova Micro, cutting inference cost by over 95% and latency by
Qwen3.6 GGUF Benchmarks
Un sloth claims top KLD-vs-disk-space performance for Qwen3.6-35B-A3B quants in 21 of 22 pareto frontier comparisons.
From hours to minutes: How Agentic AI gave marketers time back for what matters
AWS Marketing and Gradial used Amazon Bedrock to cut page assembly from 4 hours to ~10 minutes.
AWS Nova Forge SDK Tutorial: Fine-Tune Nova Models With Data Mixing
AWS publ ishes step-by-step Nova Forge SDK guide; data mixing yielded 12-point F1 gain while preserving MMLU baseline scores.
Build a Secure, Always-On Local AI Agent with OpenClaw and NVIDIA NemoClaw
NVIDIA's NemoClaw and OpenClaw framework let developers run persistent, secure AI agents locally without cloud dependency.
Qwen 3.6 is the first local model that actually feels worth the effort for me
Alibaba's Qwen3.6 35B-A3B runs Q8 at 170 tokens/ sec with full 260K context on dual consumer GPUs.
Move to local models
Source article is a personal support question, not a reportable AI news event.
Nous Research Open-Sources Her mes Agent, a Self-Improving AI Agent Framework
Hermes Agent hits 90 K+ GitHub stars with persistent skill memory and three-layer architecture across 200+ models.
Opus 4.7 来了，我并不建议你升级
Anthrop ic's Opus 4.7 removes temperature/top_p/top_k controls and inflates token counts by up to 1.35x.
Systematic Debugging Guide: A Detective Framework for Root Cause Analysis
A Chinese developer tutorial outlines a four-phase systematic debugging methodology replacing ad-hoc fixes.
Anthropic's 1M Context in Claude Code: Session Management Is the Real Story
Anthropic's official Claude Code guidance re frames 1M context as a session discipline problem, not a capacity win.
Claude Opus 4.7 Launches: 64.3% S WE-Bench Score, Higher Image Resolution
Anthropic ships Claude Opus 4.7 with self-verification coding, 2,576px image support, and no price increase.
Anthropic Adds ID Verification to Claude, Blocking Chinese Users
Anthropic's new real -time ID and facial verification system effectively bars Chinese mainland users from Claude access .
Lalamove Cuts Translation Costs 90% With 3-Agent LLM Pipeline
Lalamove deployed a three-agent LLM framework — translation, QA scoring , and compliance — slashing localization costs by 90% and reducing turnaround
Qwen3.6-35B is worse at tool use and reasoning loops than 3.5?
Community testers report Qwen3.6-35B enters infinite reasoning loops more than Qwen3.5 on agentic coding tasks.
PSA: Qwen3.6 ships with preserve_thinking. Make sure you have it on.
Qwen3.6 introduces preserve _thinking flag to keep reasoning context in-context, fixing KV cache invalidation.
GPoUr with ~12gb vram and a 3080 getting 40tg/s on qwen3.6 35BA3B w/ 260k ctx
A llama.cpp fork with turbo3 KV cache quantization achieves ~40 tok/s on Qwen3-35 B-A3B with only 12GB VRAM.
Cost-efficient custom text-to-SQL using Amazon Nova Micro and Amazon Bedrock on-demand inference
AWS details LoRA fine-tuning of Nova Micro for custom SQL dialects, hitting $0.80/month at 22,000 queries via serverless inference.
DeepMind’s New AI: A Gift To Humanity
Google DeepMind has released Gemma 4, a new family of open-weight models available under the Apache 2.0 license.
Meta's AI Agents Recover Hundreds of Megawatts by Automating Infrastructure Efficiency
Meta's unified AI agent platform compresses 10-hour manual regression investigations to 30 minutes, recovering hundreds of megawatts fleet -wide.
Deploy Postgres and MySQL databases with PlanetScale + Workers
Cloudflare Workers gains native PlanetScale Postgres and MySQL provisioning via dashboard, with unified billing launching next month.
Robots Are Finally Starting to Work
Physical Intelligence is training a single foundation model to control multiple robot platforms zero-shot, skipping per-task data collection.
How to Build Vision AI Pipelines Using DeepStream Coding Agents
NVIDIA DeepStream 9 integrates with coding agents like Claude Code and Cursor to auto-generate real-time vision AI pipeline code.
Alibaba Releases Qwen3.6-35B-A3B Mixture-of-Experts Model
Alibaba's Qwen team releases Qwen3.6-35B-A3B, a 35B-parameter MoE model activating 3B parameters per token.
Qwen3.6-35B-A3B released!
Alibaba's Qwen team releases a 35B sparse MoE model with only 3B active params under Apache 2.0.
Cloudflare’s AI Platform: an inference layer designed for agents
Cloudflare's AI Platform now routes 70+ models from 12+ providers via one API endpoint and shared credits.
Building the foundation for running extra-large language models
Cloudflare details prefill- decode disaggregation and hardware configs powering Kimi K2.5 on Workers AI, achieving 3x speed gains.
OpenCLI Turns Any Website Into a Zero-Cost CLI Agent Tool
OpenCLI generates deterministic JS adapters once via LLM, then executes them zero-cost — 15.6k GitHub stars.
Speculative Decoding on AWS Trainium2 Cuts LLM Lat ency Up to 3x
AWS benchmarks show speculative decoding with vLLM on Trainium2 reduces inter -token latency up to 3x for decode-heavy workloads.
Gemma 4 and Qwen 3.5 GGUFs: Detailed Analysis by oobabooga
Oobabooga published 5 benchmark reports covering 70-90 GGUF quants each for Gemma 4 and Qwen 3.5 models using KL Divergence methodology.
Gemma 4 Jailbreak System Prompt
A system prompt designed to bypass Gemma 4's safety filters is circulating on Reddit with 112 upvotes.
Hermes Agent Framework Hits 85K Stars With Self-Evolving Memory
Nous Research's Hermes Agent, open-sourced in February 2026 , reaches 85K GitHub stars with a four-layer memory architecture and runtime skill accumu
OpenAI Launches GPT-5.4-Cyber for Vetted Security Defenders
OpenAI releases GPT-5.4-Cyber, a fine -tuned security model, to verified defenders via its Trusted Access for Cyber program .
Local AI is the best
A Reddit post praising local AI tools contains no verifiable news, data, or technical developments.
Claude Code Desktop Rebuilt Around Parallel Agent Execution
Anthropic redesigned Claude Code desktop from scratch to run multiple AI coding agents simultaneously.
AI 自动值夜班时代来了！Claude Code 刚刚推出 Routines
Anthropic releases Claude Code Routines in research preview, enabling scheduled and event-driven autonomous coding tasks on Anthropic's cloud infrast
How Guidesly built AI-generated trip reports for outdoor guides on AWS
Guidesly's Jack AI uses AWS Lambda, Step Functions, and Amazon Bedrock to auto -publish trip content after each outdoor guide booking.
Best practices to run inference on Amazon SageMaker HyperPod
AWS details H yperPod inference deployment patterns, claiming up to 40% total cost of ownership reduction for GPU work loads.
AWS Adds Use-Case Deployment Presets to SageMaker Jump Start
SageMaker JumpStart now offers task -aware deployment configs optimized for cost, throughput, or latency by use case.
Alibaba Cloud PAI Processes 2M Videos in 200 Min via DataJuicer
Alibaba Cloud PAI ran a 7-stage video ML pipeline on 2M files (30 K hours) across 45 NVIDIA 5090 nodes in 200 minutes.
Qwen3.5-9B GGUF Quant Rankings: Q8_0 Dominates KLD Scores
KLD benchmarks across community GGUF quants show Q8_0 variants cluster near 0.001 KLD, with quality degrading shar ply below Q5.
LangChain's 10 Core Modules for Agent Dev: Code Comparisons
LangChain abstracts 10 engineering layers for AI agents, from multi-vendor LLM calls to RAG pipelines and observability.
YOLOv8 Hits 111 FPS on RK3588 for Drone Power Line Inspection
Chiba University achieves 111.3 FPS on a 6 TOPS edge chip via model pruning and async NP U scheduling.
端侧AI 模型部署实战五(Android大模型加载)
Step-by-step JNI bridge implementation for running quantized LLMs on Android using llama.cpp.
Claude Code Skills vs MCP: Architecture Deep Dive
Anthropic's Claude Code uses two distinct extension layers: Skills for reusable domain workflows and MCP for real-world tool connectivity.
Component Reuse Enforcement via AGENTS.md, Hooks, and Skills
Dew u Engineering built a three-layer AI skill system to enforce component reuse before new component creation.
Claude Teammate Mode: Multi-Agent Game Dev Post mortem
Developer deploys Claude's experimental multi-agent Teammate mode to build a TCM learning game — and documents where the workflow breaks down.
Claude Code MCP Plugin Architecture: Cross -Process Tool Proxy Dissected
Source analysis reveals Claude Code uses stdio- based MCP servers as isolated subprocesses, proxying external tools transparently to the model.