Alibaba Debuts Qwen 3.8, Netflix Opens Up Its In-House LLM Infrastructure, Claude Code Drops Weekly Limits — AI News Roundup for July 19, 2026

Alibaba Releases "Qwen 3.8" — 2.4 Trillion Parameters, Claiming "Second Only to Fable 5"

On July 19, Alibaba's Qwen team announced "Qwen 3.8," a multimodal AI model with 2.4 trillion parameters. A preview version is currently available, and the Qwen team claims the model's performance is on par with leading models, at a level "second only to Fable 5." The release is widely seen as a direct answer to competing models such as Moonshot's Kimi K3.

The same day, Reddit's LocalLLaMA community was abuzz over an OSS-related meetup held in Shanghai, and with Qwen 3.8 launching right on its heels, some voiced hopes that other players would follow with competing releases. The model appears set to be offered with open weights, so its impact on local and research use cases is drawing attention as well.

Netflix Shares the Technical Details of Its "In-House LLM Inference"

Netflix's tech blog published an article detailing how the company builds and operates its LLM inference infrastructure in-house. The know-how behind running large-scale inference on their own stack without relying on external APIs makes for a case study that only a major platform leveraging LLMs across diverse use cases — from content recommendations to production support — could offer.

The post was also picked up on Hacker News, but with just 1 point, the discussion hasn't gained much traction yet. That said, sharing insights grounded in real-world operations carries significant value from an engineering perspective, and it will be worth watching how the conversation develops.

Anti-AI Protest at OpenAI HQ, Body Bags Placed Outside

Protesters descended on OpenAI's San Francisco headquarters, staging a forceful demonstration that lined body bags up outside the office. It was a symbolic event in which concerns over the risks AI technology poses to society became visible as concrete protest action.

The story was covered by mainstream media and distributed to Japanese-language audiences as well, showing that backlash against AI continues not only on the tech side but as an ongoing social phenomenon.

Claude Code Drops Its Weekly Limits for a Limited "May–August 2026" Window; 5 Free AI Coding Tools Also Draw Attention

A Qiita article reported that Anthropic is running a campaign that removes Claude Code's weekly usage limits for a limited period from May to August 2026. The same article also spotlights five AI coding tools that are free or low-cost, breaking down the characteristics of each and how to choose between them.

Companies that have adopted Claude Code in earnest are reportedly seeing a growing number of engineers who "barely write any code themselves," so the removal of the limits looks set to have a direct impact on Japanese development teams.

"Codex 5.6 Sol Has Surpassed Fable" — A Firsthand Take on What's Different

Also on Qiita, an author analyzed the reasons they felt Codex 5.6 Sol has "surpassed Fable," based on hands-on experience. The author points to a major difference: instead of asking the AI to "write code," they now hand it the whole job with "get it to this state." They liken the gap to the difference between having someone write you a recipe and actually having the dish served at your table, suggesting that the agentic nature of AI may have improved qualitatively.