Kimi K3 is now part of PatentBench.
Results for ChatGPT 5.6 Sol, Claude Opus 4.8 and DeepSeek V4 Flash are updated. Same 340 patent-family samples. Patsnap Eureka still leads: - 85% X Hit Rate - 37% X Recall Rate
Maxymize AI digest • 4 sep
The frontier just got a full-number drop and a $13B open-source plot twist. Coffee ready? AI brief ㅤ • GPT-6 Astra rolling out – 99.9% ARC-AGI-3 (adapter), critical cyber tier, +70% token efficiency, $10/$50 ㅤ • NVIDIA locks Hugging Face
Pretty impressed to see lfm2.5-vl by liquidai complete the task accurately too
While being noticeably faster than deepseek v4 flash vision and glm 5.3 flash I honestly didn’t expect a 3b model to do this well on a robotics task.
Inspired by Sentdex post, I was curious to try the same idea and see how far general multimodal models can get in a robotics environment.
This was my first time trying something like this, so just a basic test: DeepSeek V4 Flash Vision vs GLM 5.3 Flash on a simulated Franka
You have reached the end of the archive
All of deepseek