#Qwen
Alibaba представила QwenGyre: решение проблемы простоя GPU при обучении сложных ИИ-агентов
Новый фреймворк QwenGyre от исследователей Alibaba, Университета Цинхуа и USTC устраняет вычислительные задержки при обучении автономных агентов на сверхдлинных задачах, ускоряя процесс почти в два раза.
Thomson Reuters Spent 40 Million Dollars on a Model That Surpassed Claude Sonnet
Large media holding Thomson Reuters built its own AI model Thomson based on Qwen, which scored 78.5% on aggregated benchmarks and surpassed Claude Sonnet 5. Development took 3 months and 368 GPU B200, total cost — 40 million dollars.
The Real Speed of Speculative Decoding: Why LLMs Don't Accelerate as Promised in Papers
The 6x inference speedup claimed in research papers often turns into an actual slowdown in practice. We explore how speculative decoding really works and what drives real-world performance gains.
Experiment: Can AI Models Catch All Bugs and Merge External Code into a Flawless Solution?
A practical test demonstrated why developers shouldn't blindly trust a single language model: four LLMs searched for seven bugs in a script, while seven others attempted to assemble a flawless version from external fixes.
AI Price War: How Open Models and Cheap Inference Are Erasing Giants' Superprofits
The artificial intelligence market has been gripped by fierce price competition. Open architectures like Qwen and GLM, forced price cuts by OpenAI, and in-house hardware development are transforming access to neural networks from a premium service into a basic utility.
Testing Qwen3.8-Flash-Next: Comparison with Production Models and the Importance of Multiple Runs
Recent benchmark tests of the Qwen3.8-Flash-Next architecture showed quality on par with Qwen3.8-27B alongside notable architectural innovations, though production adoption remains hindered by limited inference engine support.
Qwen 4 Architectural Breakthrough: Inside the Qwen3.8-Flash-Next Model
Alibaba has introduced Qwen3.8-Flash-Next, an architectural preview of the upcoming Qwen 4. The 180-billion-parameter model activates just 6 billion parameters per token and can run on 128 GB of system memory.
Digital Time Machine: How to Animate Old Photos and Transport Friends into Them Using AI
A practical guide to merging archival photographs with modern portraits into realistic videos using Minimax H3, Qwen Edit, and auxiliary AI tools.
SpaceX and Nvidia Join Forces for Orbital AI Data Centers Targeted for 2027
SpaceX has partnered with Nvidia to power its orbital Starmind data centers with custom Vera Rubin hardware, aiming for launch by late 2027 amid growing ground-based infrastructure constraints.