SGLang

Advances in image composition are simultaneously broken by Reve 2 and Ideogram 4, with Ideogram 4 now the top-ranked open image model on Arena. Microsoft released MAI-Thinking-1 achieving 97% on AIME 2025 without synthetic data or distillation, publishing detailed training stacks and MoE scaling. Frontier Tuning enables enterprise workflow models to reach GPT-5.4 quality with up to 10× efficiency gains, while Gemma 4 12B and others strengthen local-first deployment momentum.

Together Inference Engine delivers 31% more TPS than next fastest OSS engine on same hardware, maintains 2× better TTFT at saturation. Performance gains come from full-stack optimization.

Chinese GPU manufacturer Moore Threads attracts core developers of multiple large model inference frameworks through open source activities.

DeepSeekV4发布，通过组合创新和工程优化，在R1的“测试时扩展”范式下，实现百万上下文从理论到实用的飞跃，对Agent和多步复杂任务具有重要意义。

NVIDIA AI 报告称，SGLang 在 Blackwell 硬件上使用 DeepSeek-V4 模型解码达到 180 tok/s/GPU 的速度，约 1M 上下文，得益于 LMSYS 组织针对 Blackwell 的特定优化，提高了混合稀疏注意力的利用效率。

Google announces its model weights are compatible with major open-source ecosystems and can be directly downloaded from Hugging Face and Kaggle, lowering deployment barriers.