Claude Opus 4.5 Tops SWE-Bench as DeepSeek’s Engram & Prompt Hacks Redefine LLM Optimization
AI Research & Model Advancements
DeepSeek Engram: A Static Memory Unit for LLMs
DeepSeek AI introduced Engram, a static memory module for LLMs that avoids recomputing static knowledge, improving reasoning, math, and code performance. It enables massive memory scaling without GPU constraints and frees attention for global reasoning. The paper and a video explanation are available on GitHub.
Prompt Repetition Boosts Non-Reasoning LLM Performance
A new arXiv paper reveals that repeating prompts improves accuracy in non-reasoning LLMs (e.g., data retrieval) without increasing latency. Deepseek models are highlighted as key beneficiaries of this technique.
Benchmarking & Model Performance
December 2025 SWE-bench Leaderboard Update
Nebius updated the SWE-bench leaderboard with December 2025 results on 48 fresh GitHub PR tasks. Claude Opus 4.5 leads (63.3% resolved), followed by GPT-5.2 (xhigh effort, 61.5%). GLM-4.7 is the top open-source model, while Gemini 3 Flash Preview outperforms Gemini 3 Pro Preview.
Industry Partnerships & Hardware
OpenAI-Cerebras Deal to Accelerate Codex Inference
OpenAI partnered with Cerebras Systems to significantly improve Codex inference speeds, aiming to enhance AI-driven coding workflows.
US Eases Export Controls: Nvidia GH200 & AMD Mi325X Shippable to China
The U.S. amended export restrictions, allowing Nvidia GH200 and AMD Mi325X chips to be shipped to China, potentially boosting regional AI hardware availability.