Category: AI Research
-

China Just Open-Sourced a 1.6 Trillion Parameter Model
Meituan, the Chinese food delivery giant, just released LongCat 2.0. It is a 1.6 trillion parameter model trained from scratch on a 50,000-GPU cluster built inside China, and they open-sourced it. Here is what that means: China is no longer waiting for permission to build frontier-scale AI. They are building it anyway, using domestic silicon…
-

Google Just Dropped the Model That Beats Everyone on Science
Google released Gemini 2.5 Pro with Deep Think on June 22, and the benchmarks are not close anymore. They are not even in the same building. Here are the numbers: 89.8 percent on MMLU-Pro. 82.4 percent on GPQA Diamond. GPQA is the science one: the test where you need actual knowledge, not pattern-matching. Gemini is…
-

OpenAI’s Model Just Diagnosed 18 Kids Doctors Couldn’t
Rare disease diagnosis is a game of connect-the-dots played with incomplete information. Doctors see symptoms. They run tests. Sometimes, even after months of work, they run out of leads. Then they get stuck. A team at Boston Children’s Hospital decided to try something different. They took 376 cases that had been analyzed multiple times by…
-

Google’s AI Just Quietly Fixed Something Nobody Else Could
AlphaEvolve is not new. Google DeepMind showed it off a while ago. But the new part is what it is doing now, out in the world, where it actually matters. This coding agent-this thing that writes and rewrites algorithms-started as a research project. Now it is in production. It is making real saves. Start with…
-

Asian AI Labs Are Filling Anthropic’s Export Gap
Anthropic’s export mess is already pushing other labs to move. Chinese security firm 360 has unveiled Tulongfeng, and Sakana AI has launched Fugu, both leaning hard on the same opening: if the U.S. gate gets tighter, somebody else will try to serve the market. That is why this story matters. It is not just that…
-

A 3B Model Just Outran the Megamodels at Reasoning
Weibo AI shipped VibeThinker-3B last month and it’s blowing the reasoning-to-parameters ratio to pieces. A three-billion-parameter model that scores 94.3 on AIME26, an 80.2 on LiveCodeBench v6, and passes recent unseen LeetCode contests at 96.1%. Here’s the punch: these are the benchmarks that GPT-5.5, Opus 4.8, o1, and the full-size reasoning models fight over. VibeThinker-3B…
-

Baidu’s Long-Horizon OCR Just Hit 1.8K GitHub Stars
Baidu shipped Unlimited-OCR on June 22, and it did 1.8k GitHub stars in under 24 hours because it solves a problem everyone in document AI actually has: context doesn’t stop at page breaks. Until now, OCR systems had to chunk documents, process each chunk separately, then glue the output back together. That’s clunky and fragile.…
-

GPT-5 Just Cracked a 3-Year Immune Cell Mystery
Derya Unutmaz, an immunologist at The Jackson Laboratory for Genomic Medicine, had been stuck on the same question for three years: how does glucose actually change the way T cells develop and fight cancer? That’s not a small question. T cells are the immune system’s precision weapons. Getting glucose metabolism wrong breaks the entire picture…
-

Open Source Just Won the Download Race
Something shifted in June 2026. The open-source models started winning. DeepSeek V4.1 Flash is the trending model on Hugging Face right now. Qwen is up there. Llama. Gemma. Chinese models hold five of the top ten download slots, the highest concentration on the record. One week after the frontier wave dropped, open-weight models are not…
-

Google DeepMind Funds AI Safety Research With $10 Million
Google DeepMind just announced a 10 million dollar funding call for researchers studying a problem most people haven’t thought about yet: what happens when millions of AI agents built by different companies all interact with each other. These agents will communicate, negotiate, and transact with one another. They’ll need to do all that safely and…
