qwenGPU Compiles Qwen3.5-2B into GPU Layers
MarkTechPost described the qwenGPU project, which compiles a 4-lane SIMT GPU into 45 untrained layers of the Qwen3.5-2B model. The model writes its own CUDA launch steps, using Gated DeltaNet state as registers and memory.
Summarized by AI from one outlet
Source
- MarkTechPostRead original ↗
Latest
- Models
OpenAI Launches GPT-6 Cyber, Finds 81 Unknown Firefox Flaws
OpenAI released its latest AI model, GPT-6 Cyber, which discovered 81 unknown vulnerabilities in Firefox. The move highlights AI's potential in cybersecurity and raises concerns about misuse risks and restricted distribution mechanisms.
- Models
Meta rolls out Muse Wonder models with enhanced memory and focus features
Meta announced the rollout of dedicated models trained on Muse Wonder for Muse and Meta AI users. The update brought down Active Memory and Active Focus features, significantly improving task completion rates and reducing information forgetting, enhancing AI assistant performance in complex scenarios.
- Models
Google releases Gemini Omni 1.2 Flash for 15-second video generation
Google DeepMind released the Gemini Omni 1.2 Flash model, extending native video generation from 10 to 15 seconds and introducing multi-turn conversational editing. The model also cuts API pricing by 30%, reducing the cost per second at 720p to approximately $0.07 to lower marginal costs for video generation.
- Models
Zhejiang PhD releases multi-physics world model
A post-2000 PhD student at Zhejiang University released a new model that integrates over a dozen physical elements. Developed on the Huawei Ascend native ecosystem, it marks progress in China's version of Cosmos.
All news → · All rights belong to the original creators