DeepSeek V4 Is China's Strongest, but Still 8 Months Behind the US Frontier — CAISI Evaluation Report

CAISI, an institute under the U.S. National Institute of Standards and Technology (NIST), has released an evaluation report on DeepSeek V4. It concludes that while DeepSeek V4 has become the most powerful AI model in China, it still lags roughly eight months behind U.S. frontier models.

The assessment is based on systematic benchmarking by an independent organization, which is significant in that it objectively establishes DeepSeek V4's standing — a question previously debated mostly within the community. It also reconfirms that while Chinese AI development is rapidly closing the gap, a meaningful distance remains at the cutting edge.

ByteDance's Drug Discovery Arm Presents AI-Designed Therapeutics at International Conferences

The drug discovery division of ByteDance, TikTok's parent company, has presented research on AI-designed therapeutics at multiple international conferences. According to a report by the South China Morning Post (SCMP), this is a notable development signaling that AI is starting to deliver practical results in drug discovery.

AI-driven drug discovery has been led mainly by Western pharmaceutical companies and startups, and the entry of Chinese tech firms like ByteDance suggests that competition in this field could intensify further.

Suno Signs AI Music Licensing Deal with Warner Music and Acquires Songkick

Suno, an AI music generation company, has signed a licensing agreement with Warner Music Group. As part of the deal, it was also revealed that Suno will acquire Songkick, the concert information platform.

Suno has a history of facing copyright infringement lawsuits, making this partnership with a major label a significant turning point for the AI music industry. As consensus builds with rights holders, the framework for the legality and business models of AI-generated music is gradually taking shape.

Qwen3.6-27B vs Coder-Next — Real-World Benchmarks Reveal Sharp Differences

A benchmark has been making waves in the LocalLLaMA community, putting Qwen3.6-27B and Coder-Next through a thorough head-to-head comparison over 20 hours on two RTX PRO 6000 Blackwell GPUs.

On overall scores, the two models were nearly evenly matched, with no statistically significant difference. The results, however, revealed interesting strengths and weaknesses depending on the nature of the task. Coder-Next posted a 10/10 success rate on bounded business document generation and document consolidation tasks, and can run at 60–100x lower cost than the 27B. On live market research tasks, by contrast, Coder-Next collapsed to 0/10 while the 27B scored 8/10.

Another interesting finding: disabling thinking mode on the 27B (`--no-think`) yielded the most stable task completion, hitting a 95.8% success rate. This suggests that the verbosity of the reasoning process can, paradoxically, trigger loops.

The AI Spending Trap — Adoption Advances While Outcomes Lag Behind

An Age of Product article, "The AI Spending Trap: Why Adoption Outpaces Outcomes," is drawing attention. It argues that while enterprise AI adoption is advancing rapidly, actual business outcomes often fall short of the level of investment.

The article identifies structural problems: AI spending continues to balloon, yet actual ROI (return on investment) remains difficult to measure, and organizational transformation processes have not kept pace with technological progress.

---