AI Briefing#llm-efficiency
The latest AI and developer news about #llm-efficiency, with the original source and a short summary.
The latest AI and developer news about #llm-efficiency, with the original source and a short summary.
The method uses second-order interactions to maintain performance during aggressive expert removal, outperforming first-order baselines on Qwen3.5 and GLM-4.5 models.
The method achieves a 27.8% increase in mean acceptance length over DFlash on Qwen3-8B at temperature zero.
The method uses DPO on self-generated traces to teach models when to stop reasoning, improving the accuracy-efficiency trade-off across math and code benchmarks.
Trajectory introduces Intelligence Density and density-aware training, demonstrating significant reductions in output tokens and improved efficiency on legal and financial benchmarks without sacrificing accuracy.
Fireworks Research releases Ember-1, a specialized model built on Kimi K3 that delivers comparable performance with 40% fewer tokens, now available as a Research Preview.
During the third week of September 2026, papers on pre-compilation and distillation paradigms, such as 'Compile by Training', 'R4T', and 'Breaking the Token Ceiling', gained attention for reducing inference costs.
**Toast 1**, a specialized search agent that dramatically improves cost and speed while maintaining search performance, has been released.
MoBA, a technology that applies the MoE principle to attention to maximize the efficiency of long-context processing, has been announced.
This article introduces engineering optimization techniques that leverage the Hybrid SWA architecture of the MiMo-V2.5 model to reduce KVCache storage space and compute to about 1/7 the level.
By recycling information from discarded tokens through RCD technology, the accuracy and inference speed of diffusion language models have been dramatically improved.
DeepSeek has open-sourced an optimization technique that maximizes inference efficiency, boosting generation speed by up to 85%.
An Abstract CoT technique has been proposed that allows LLMs to reason efficiently using meaningless tokens.
By leveraging high-quality reasoning data, a 15B reasoning model was distilled into a 2.1x faster Mamba hybrid without performance degradation.