Tech Times

Compile Once, Run Offline: New AI Method Matches 32B Models With a 23MB File

Local AI inference at 32B-parameter quality, no cloud API required: University of Waterloo researchers released PAW on July 2 ...

Waterloo's PAW compiles task specs into 23MB LoRA adapters a 600M-parameter model runs entirely offline.

Local AI inference at 32B-parameter quality, no cloud API required: University of Waterloo researchers released PAW on July 2, 2026, a system that compiles any natural-language task spec into a 23MB ...

Fable 5 Breach Leaks Cryptic AI Chain of Thought Shorthand

Fable 5's chain of thought has leaked, showing math-like shorthand, while its three-layer defense classifiers block most jailbreak attempts.

DeepSeek open sources DSpark, a new framework to speed up LLM inference by up to 85%

DSpark can make decoding faster, but acceptance quality still determines how much speed the system actually realizes.
GitHub

DPad: Efficient Diffusion Language Models with Suffix Dropout

Efficiency: DPad-enhanced dLLMs achieve up to a 61.39× speedup over vanilla dLLM baselines. Accuracy: DPad-enhanced dLLMs achieve up to a +26.46% improvement over vanilla dLLM baselines. (Evaluation ...
GitHub

DyCon: Dynamic Reasoning Control via Evolving Difficulty Modeling

Difficulty modeling should align with trajectories rather than solely the initial question. Existing methods typically rely on static difficulty estimations, handcrafted confidence heuristics, or ...