Gemma 4: Google Finally Says "Open Source" and Means It
Google released Gemma 4 this week under the Apache 2.0 license. That matters more than any benchmark number. Previous Gemma models were open-weight but not truly open-source. You could look but not freely build on top of them for commercial use without navigating restrictions.
Gemma 4 comes in four sizes from 2B to 31B parameters. All of them handle video, images, and text natively. The smaller models run on Android phones and laptop GPUs. The 26B model uses a mixture-of-experts setup with only 4B active parameters, so it punches way above its weight class while staying efficient.
For builders shipping products, this is the headline. You get a multimodal model trained on 140+ languages with a real open-source license. No lawyers needed. Download it from Hugging Face or Ollama and start building today.
Qwen 3.6-Plus: A Million Tokens for Agentic Coding
Alibaba dropped Qwen 3.6-Plus with a 1,000,000-token context window. That is roughly 2,000 pages of text in a single prompt. The model is specifically built for agentic coding, meaning it can break down programming tasks, write code, test it, and fix bugs in a loop.
The benchmarks are hard to ignore. Qwen 3.6-Plus scored 61.6 on Terminal-Bench 2.0, beating Claude 4.5 Opus, and 78.8 on SWE-bench Verified. Early testing on OpenRouter shows it running 2-3x faster than Claude Opus in tokens per second. And right now, the preview is free on OpenRouter.
If you are doing any kind of repo-level code analysis or building coding agents, this is worth testing today. A million tokens means you can feed it an entire codebase without chunking. That changes the workflow completely.
Conway: Anthropic Tests an Always-On Agent
Reports surfaced this week that Anthropic is internally testing something called Conway. It is a persistent agent platform that keeps Claude running as a background process on your system. Think of it as Claude living in a sidebar, 24/7, responding to webhooks and handling multi-step tasks without you opening a chat window.
The leaked details show an extensions system with a .cnw.zip package format, webhook URLs that wake the instance when external services call them, and Chrome browser integration. This points toward a future where your AI assistant is not something you talk to. It is something that works for you while you sleep.
No release date yet. But the architecture suggests Anthropic is serious about competing in the autonomous agent space that OpenAI has been pushing with GPT-5.4's computer-use features.
What This Means for Builders
The pattern is clear. Models are getting cheaper, more capable, and more autonomous every week. Google giving away a strong multimodal model under Apache 2.0 puts pressure on every paid API. Alibaba offering a million-token coding model for free on OpenRouter puts pressure on every coding tool. And Anthropic building always-on agents puts pressure on every workflow that still requires a human to type "go."
For small shops and solo developers, this is the best time to be building with AI. The tools are better and cheaper than they were even a month ago. Pick one of these and ship something this weekend.