AI Research Brief

Mistral has released Mistral Large 4

OpenAI models are coming to Google Cloud

The authors propose LongZero, a method for enhancing long-context reasoning in LLMs via reinforcement learning from zero external data. By having the model synthesize its own challenging question-answer pairs and learning from them, it bypasses the scarcity of high-quality long-context training data, offering a scalable path to improve reasoning without relying on stronger teacher models

DeLoop addresses the common failure mode of long-horizon agents getting stuck in repetitive loops by using self-supervised learning. It treats a looping action as a negative example and the subsequent non-looping action as a positive one, training via direct preference optimization without requiring task rewards or LLM judges, significantly improving agent reliability under context compression

This work introduces PhysiCalWorld, which calibrates action-conditioned world models to instance-specific physical dynamics (like mass or friction) via in-context prompting. By encoding an exploratory interaction clip into a compact physics descriptor token, it grounds predictions in actual scene behavior, reducing drift in novel environments without fine-tuning or analytical physics modeling

Recursive releases Flint, a 39.8B parameter mixture-of-experts model (2.5B active) designed for coding agents. It achieves strong results on coding benchmarks, including 89.7% on SWE-bench Pro v2 and 54.9% on Terminal-Bench 2.1, and is released under Apache 2.0 with support for multi-turn tool use and reasoning

Spark-X2.6-4B is a 4B parameter model optimized for on-device agent deployment, released under Apache 2.0. It retains the architecture of its predecessor but achieves substantial gains in coding, agent capabilities, and long-context reasoning through improved post-training, making it a practical upgrade for mobile and edge AI applications

A detailed guide outlines the VRAM, RAM, and GGUF quantization requirements for running DeepSeek V4.1 Flash locally. It covers deployment via vLLM and llama.cpp, including Engram offloading strategies, providing critical data for practitioners looking to deploy this model on consumer or edge hardware

Analysis of Huawei Ascend 950PR chips running DeepSeek models provides benchmarks and insights into non-NVIDIA inference strategies. This is crucial for understanding the performance and viability of alternative hardware ecosystems for serving large language models