™Topics
All topics →Predictions
To the Arena →All news
Newest first-
Jailbroken PS5 Becomes a Local AI Server: Qwen3.8-27B Hits ~20 tok/s, Small Models Top 100
Why it mattersThe first time a game console has run a 27B model at usable speed through open-source porting alone — a 16GB, 448GB/s-bandwidth PS5 now sits in desktop inference territory, and the community just rewrote the hardware bar for local AI.
-
Meta Releases Muse Spark Open, Teaming Up with Apple to Put an 816B-Parameter Model on Mac Studios
Why it mattersIt is Meta's second open-weight model since its AI overhaul, and a rare public team-up between two longtime rivals.
-
Apple Intelligence Launches in China in December with iOS 27.3, Excluding New Siri AI: Report
Why it mattersModels run via Alibaba's Qwen and Baidu under MIIT permits granted Sept 29, ending two years of China-side absence.
-
Alibaba's Qwen Introduces QwenGyre: Elastic RL for Extreme-Long-Horizon Agents
Why it mattersOn the flagship Qwen 3.8 2.4T, 48 training steps lifted NL2RepoBench from 52.5% to 58.5%, with up to 1.85x end-to-end training speedup.
-
Alibaba Open-Sources Its Internal AI Code Review Tool
Why it mattersIt ran inside Alibaba for two years and caught millions of defects; on its own AACR-Bench it beats general-purpose agents on precision and F1 with the same model while using about 1/9 of the tokens.
-
Qwen-Image-2.1-Turbo: High-Speed Image Generation in 8 Steps
Why it mattersIt cuts generation from dozens of steps to eight and outputs up to 2K; it ships under the Qwen Research License rather than Apache 2.0, so commercial use needs separate licensing.
-
JetBrains Releases Mellum2.1, an Open Model for Coding Agents
Why it mattersMellum2.1 keeps the 12B mixture-of-experts design with 2.5B active parameters under Apache 2.0, and was trained with millions of sandboxed runs across thousands of in-house RL environments; under heavy load it serves nearly twice the tokens of Qwen3.5-9B.
-
Open-Source Strata Runs a 125-Billion-Parameter Model on a 12GB Gaming PC
Why it mattersA 125-billion-parameter model runs on a gaming PC with 12GB of VRAM and 32GB of RAM; community tests show about 2x llama.cpp speed, versus the advertised 6x.
