Tencent выпустила новую модель Hy4, которая может составить серьёзную конкуренцию сильнейшим ИИ-моделям с открытыми весами. По результатам внутренних испытаний Tencent, Hy4 набрала 75,2 балла в бенчмарке APEX-Agents, опередив DeepSeek V4.1 Flash-1017, получившую 72,1 и 72,9 балла. Впрочем, в других испытаниях преимущество переходит от одной модели к другой. Стоимость использования Hy4 через API составляет $0,834 за миллион входных токенов и $2,501 за миллион выходных. Модель также поддерживает работу с изображениями и предназначена для решения широкого круга задач, включая программирование, финансовый анализ, создание игр и научные исследования. Кроме того, Tencent намерена опубликовать веса Hy4. https://hy.tencent.ai/research/hy4 До 15 ноября Hy4 можно бесплатно использовать в WorkBuddy. Правда, с ежедневными лимитами.
Tencent has released a new model, Hy4, which can seriously compete with the strongest open-source AI models...
Tencent has released a new model, Hy4, that can seriously compete with the strongest AI models with open weights. Based on Tencent's internal testing, Hy4 scored 75.2 in the APEX-Agents benchmark, ahead of DeepSeek V4.1 Flash-1017, which scored 72.1 and 72.9...
AI Summary & Key Takeaways
- Tencent has released a new model, Hy4, that can seriously compete with the strongest AI models with open weights
- Based on Tencent's internal testing, Hy4 scored 75.2 in the APEX-Agents benchmark, ahead of DeepSeek V4.1 Flash-1017, which scored 72.1 and 72.9...
Tencent has released a new model, Hy4, which can seriously compete with the strongest open-source AI models...
Related stories
Six plugins for DeepSeek Harness that will help reduce token consumption and bring order to the context.
Six plugins for DeepSeek Harness that will help reduce token consumption and bring order to the context. Some of them are responsible for collecting and storing information, others expand the capabilities of AI agents, and others allow you to control their work. Everyone solves their own problem, so...
OpenAI told how it tests agents on real everyday tasks of users
✔️ OpenAI described how it tests agents on real everyday tasks of users. To do this, they use the internal LifeVal benchmark, which evaluates the practical benefits of neural networks in everyday life. The creation of the OpenAI test was led by the understanding that a simple test of basic knowledge...
Subscription to AI no longer means unlimited: what has changed with ChatGPT, Claude and Gemini
Subscribing to AI no longer means unlimited: what has changed for ChatGPT, Claude and Gemini ⠀ You save queries, take a simpler model, transfer tasks between neural networks or look for workarounds. All in order not to hit the limit in the middle of work. Further restrictions will only be...
CARD: A Proof-Carrying Certificate-to-Action Interface for Relative-Measurement Graphs
arXiv:2611.03651v1 Announce Type: new Abstract: Geometric certificates make mathematical claims about specified estimator objects, while robotic action requires a physical transform, a service decision, and timely proof support. We present CARD, a proof-carrying interface that…