China's Short Drama Industry Shifts Entirely to AI Mass Production — An Astonishing Pace of 470 Titles Released Per Day

According to MIT Technology Review, China's short drama industry is rapidly shifting to mass production of AI-generated content. Short dramas are a content format made up of brief 1–2 minute episodes, characterized by dramatic storytelling optimized for smartphone viewing. In 2024, the format generated approximately $6.9 billion in revenue in the Chinese market, surpassing the country's annual box office for the first time.

What's remarkable is the pace. According to DataEye research, an average of 470 AI-generated short dramas were being released per day as of January 2026. Production timelines have shrunk from the traditional 3–4 months to under one month, with production costs reportedly cut by 80–90%. FlexTV has completely halted traditional live-action production and fully transitioned to AI generation, while Kunlun Tech (DramaWave's parent company) already offers more than 1,000 AI titles on its platform.

The transition is also having a major impact on the labor market. Roles such as camera crews, lighting technicians, and makeup artists have nearly vanished from AI production sets, replaced by a new job category: the "AI asset curator." This role involves converting scripts into prompts and generating reference images of characters and scenes for AI video models, and hundreds of such job listings have reportedly been posted on Chinese job sites.

Screenwriter Phoenix Zhu described how writing scripts for AI has come to double as the cinematographer's job: "Before AI, you could get away with 'he gave her a cold look.' Now you have to write it visually and specifically, like 'a cold ray of light shot from his eyes.'"

For the industry as a whole, Omdia estimates that the global micro-drama market reached $11 billion in 2025 and is projected to grow to $14 billion by the end of 2026.

Intern-S2-Preview: A 35B-Parameter Multimodal Model Specialized for Scientific Tasks

Intern-S2-Preview, which has been making waves in Reddit's LocalLLaMA community, is a 35B-parameter science-specialized multimodal foundation model released by InternLM. It continues pretraining on top of Qwen3.5 and pursues "task scaling" in addition to conventional parameter and data scaling.

Specifically, it employs methods that increase the difficulty, diversity, and coverage of scientific tasks, scaling hundreds of specialized scientific tasks across the full chain from pretraining to reinforcement learning. As a result, it reportedly matches the performance of the trillion-scale Intern-S1-Pro on multiple specialized scientific tasks despite its relatively compact 35B parameter count.

A particularly notable feature is that by enhancing the spatial modeling of small-molecule structures and introducing a real-value prediction module, it has become the first open-source model to combine materials crystal structure generation capabilities with general-purpose capabilities. It also adopts weight-shared Multi-Token Prediction (MTP) and Chain-of-Thought compression techniques during reinforcement learning, improving inference efficiency as well.

UK Unveils Sovereign LLM Inference Platform relax.ai — 52-Comment Discussion on HN

Drawing attention with 59 points and 52 comments on Hacker News is relax.ai, a sovereign LLM inference platform from the UK. It offers an LLM inference service hosted within the UK that guarantees data never leaves the country.

This can be seen as part of the "sovereign AI" trend taking hold across countries. As interest grows in data sovereignty and autonomy over AI infrastructure, demand is rising for locally self-contained AI inference environments aimed at governments and public institutions. In the comment section, topics such as the definition of sovereign AI, performance and cost challenges, and the potential of open-source models were actively debated.

WhichLLM: A Tool That Recommends the Best Local LLM for Your Hardware, Ranked by Benchmarks

Also gathering 35 points on Hacker News, WhichLLM is a tool that ranks the best local LLMs by benchmark score based on a user's hardware specifications. Available on GitHub, it has been praised for offering a practical way to answer the question "which model will actually run well on my GPU" as local LLM options continue to multiply.

In the local LLM community, where new models are released almost weekly, verifying hardware compatibility and performance has become one of the biggest barriers to real-world use. The emergence of practical tools like this can be considered essential infrastructure supporting the broader adoption of local LLMs.

---

**Note:** This article is based on publicly available information from MIT Technology Review, WIRED, Hacker News, Reddit LocalLLaMA, and other sources. Please refer to each source for further details.