Did Claude Opus 4.8 Distill Alibaba's Qwen? Here's What the Evidence Says
Anthropic released Claude Opus 4.8 on May 28, 2026, and some users observed the model answering in Chinese that it was "Tongyi Qianwen," Alibaba’s Qwen family. The claim that Opus 4.8 had distilled Qwen spread across Hacker News, Reddit, and Chinese forums, but the behavior proved inconsistent across runs and languages. Analysts attribute the glitch to a Chinese‑language identity bug likely caused by training‑data contamination, prompt fragility, or third‑party proxy routing, not a true model distillation. No forensic evidence confirms Anthropic used Qwen outputs in Opus 4.8’s training.
What We Learned From 3 Million Downloads of Kilo Code
Kilo Code announced that its AI‑coding extension has surpassed three million downloads and processed more than 40 trillion tokens across VS Code, JetBrains, CLI, Slack, and Teams. The growth forced the team to evolve from a simple sidebar tool into a full‑stack,...
The AI Coding Revolution Hasn't Started Yet
At the Open Source Summit in Minneapolis, the author observed that most engineers— even senior staff—are still limited to basic autocomplete tools and have not embraced AI‑driven coding agents. Security approvals, underwhelming early experiences, and uncertainty about onboarding are the...
The Age of the Flash Model: Gemini 3.5, StepFun, DeepSeek and the Future of Agentic Engineering
Google unveiled Gemini 3.5 Flash at I/O 2026, a high‑performance model built for agentic engineering and already powering Search. The new tier costs roughly 60% more than budget‑friendly alternatives like DeepSeek V4 Flash but delivers frontier‑level scores (74.2% on PinchBench)...
AI Consulting Wins When It Embraces Model Freedom
The AI frontier is shifting from pure model releases to full‑scale consulting services. OpenAI is forming a $14 billion "Deployment Company" with private‑equity partners, while Anthropic has launched an enterprise‑services venture backed by Blackstone, Goldman Sachs and Hellman & Friedman. Google is supporting...
PinchBench 2.0 Is Here
PinchBench 2.0 launches as the new reference benchmark for AI coding agents, expanding from 23 to 148 tasks that mirror real‑world data analysis, DevOps, and research workflows. The update fixes scoring loopholes by normalizing results across task count and introduces...
One Company’s Blueprint for Taking a Whole Engineering Org Agentic
Ida Infront, a Swedish government‑software vendor with a 25‑year‑old platform, aims to double its development speed by year‑end. To achieve this, it has adopted Kilo Code, an agentic engineering platform, rolling it out to three pilot teams and 70 seats....
Hermes Vs. OpenClaw - When to Reach for Which Agent
The blog compares Hermes Agent and OpenClaw, two open‑source AI assistants that connect to chat apps and run tools. Hermes, launched in February 2026, centers on a learning loop that auto‑generates and refines skills, offering five sandbox backends and sub‑agent...
Mistral Medium 3.5 Is Live in Kilo Code
Mistral announced the public preview of its Medium 3.5 model, a 128‑billion‑parameter dense LLM with a 256k token context window, now live on the Kilo platform. The blended architecture unifies instruction‑following, reasoning and code generation, achieving a 77.6% score on SWE‑Bench,...
We're Back on Product Hunt — Here's How Far Kilo Has Come
Kilo Code returned to Product Hunt with a dramatically upgraded AI‑powered development suite. The new VS Code extension, part of a broader ecosystem that includes JetBrains, CLI, Cloud Agents, Slack, and a code reviewer, now boasts over 2.3 million installs. It introduces...
KiloClaw in VS Code, Kilo CLI in KiloClaw
Kilo released a native KiloClaw chat panel in the VS Code extension (v7.2.20) and added org‑aware /kiloclaw support to the built‑in Kilo CLI. The new panel lets developers converse with a persistent KiloClaw agent while coding with Kilo Code, and the...
The New AI Problem Is a Lack of New Data
The AI coding market is heating up as firms like SpaceX and OpenAI vie for high‑value developer data rather than subscription revenue. Researchers estimate the public internet’s 300 trillion tokens could be exhausted by 2026, forcing labs to turn to new...
Why Open Source Will (Still) Win in the Age of Agents
The AI tooling market is seeing closed‑source agents like Cursor and Windsurf absorbed or dissected, underscoring the risk of vendor lock‑in. OpenClaw, an open‑source agent framework launched in November 2025, quickly gathered 360,000 GitHub stars and 44,000 community‑built skills. The...
DeepSeek-V4 Is Here. So Is Everybody Else.
DeepSeek released two new models—V4‑Pro (1.6 trillion total, 49 billion active) and V4‑Flash (284 billion total, 13 billion active)—instead of a single V4 offering. Pricing for V4‑Pro is markedly higher than the disruptive rates of V3, pushing effective costs to about $1.74 per million...
Shell Security Plugin
The new Shell Security plugin links OpenClaw’s built‑in security audit with KiloCode’s Security Advisor API, turning raw JSON findings into a prioritized, plain‑language remediation report delivered inside chat platforms like Slack or Telegram. It runs the audit locally, sends only...