A new study from arXiv (2608.18230) introduces MixerLoop, a looped language model architecture that repeats only the Gated DeltaNet mixer while applying the dense feed-forward network once, and finds …
Alibaba's Qwen team released the open weights for Qwen3.8-2.4T-A95B, a 2.4T-parameter MoE flagship, on August 13, 2026, and Fixstars deployed it on a single 8x NVIDIA B300 node, achieving twice the th…
Moonshot AI released Kimi K3, an open frontier model that swept leaderboards, featuring a hybrid attention mechanism with Kimi Delta Attention (KDA), a linear attention layer derived from DeltaNet and…
Researchers at Moonshot AI introduced Kimi Linear, a hybrid linear attention architecture that outperforms full attention across short-context, long-context, and reinforcement learning scaling regimes…
A developer running a custom fork of rapid-mlx called qMLX on a 96GB Mac Studio has removed hot caching to enable multiplexing for agentic workflows, allowing concurrent sessions by relying on SSD-bas…
Alibaba's Qwen team released Qwen-AgentWorld-35B-A3B, a language world model that predicts environment states rather than actions, trained on over 10 million interaction trajectories across seven agen…
Researchers introduced Gated DeltaNet-2, a linear attention model that decouples the erase and write operations in recurrent state updates using separate channel-wise gates. The model outperforms Mamb…
Alibaba's Qwen3.6-27B dense model outperformed its own 397-billion-parameter MoE predecessor on SWE-bench Verified (77.2% vs. ~76.2%) while running on a single RTX 4090, challenging the prevailing MoE…