Latest

The non-obvious point is that YOCO-style self-forecasting caches, previously requiring dedicated training like DeepSeek-V4.1-Flash, can now be retrofitted onto off-the-shelf checkpoints with no training and no dropped tokens. Engineers deploying long-context inference in llama.cpp or vLLM can use it to cut time to first token on a single A100. Benchmarks like RULER and LongBench v2 let readers judge the quality cost themselves.

The clearest lesson is that most of the gain for the 27B model came from the serving-side inference procedure rather than fine-tuning data, which changes how practitioners should budget effort on reasoning benchmarks. Builders running small-model inference on constrained hardware can reuse the expert-pruning and KV-cache techniques directly. The study's 189-run methodology on the ARC Prize Foundation's semi-private set also offers a rare reproducible comparison for offline ARC-AGI-2 solvers.

Memory write-injection looks like a cheap way to add knowledge to a deployed model, but these results show the added facts cannot be chained into multi-hop reasoning by the model itself — orchestration must live in the application layer. Builders wiring RAG-style memory into agents should chain written atoms explicitly in prompts or system logic, and benchmark vendors should adopt coverage, synergy and distortion reporting. The tested beds include Chinese frontier models Qwen3 and DeepSeek, making the findings broadly relevant across model families.