2026-03-22 Digest

Tracked 125 · Curated 6

#1 Safely Deploying ML Models to Production: Four Controlled Strategies

Deploying machine learning models directly into production is inherently risky due to real-world complexities. This article details four controlled deployment strategies to mitigate potential degradation: A/B testing, which compares traffic between models; Canary testing, which involves a gradual rollout to a subset of users; Interleaved testing, which mixes model outputs within a single interaction; and Shadow testing. These methods allow ML teams to validate candidate models under real-world conditions, minimizing disruptions while ensuring performance metrics meet production standards before full-scale implementation.

7.2

#2 Moonshot AI Releases Attention Residuals

Moonshot AI has open-sourced its Attention Residuals project. This technical initiative focuses on optimizing residual connections within attention mechanisms to enhance the training efficiency and performance of Large Language Models (LLMs), offering developers a new approach to improving model architecture.

7.1

#3 purl: A curl-esque CLI for making HTTP requests that require payment

purl is a curl-inspired command-line tool designed for making HTTP requests that require payment. It streamlines the process for developers testing or interacting with paid API resources by integrating payment logic directly into the CLI workflow, offering a specialized solution for accessing monetized services.

6.5

#4 Introducing Single-tenant Cloud HSM

The cloud provider has launched a new Single-tenant Cloud HSM service. It provides a dedicated, highly-available cluster of HSM partitions, allowing users to retain full control over their cryptographic keys. This service is designed to enhance data encryption capabilities and support stringent security requirements.

6.2

#5 Mole Releases v1.31.0 Makima

Mole has officially released version 1.31.0, codenamed "Makima." This update delivers improved performance and more accurate system status reporting. The project remains fully open source and free, with the development team highlighting the significant contributions from the community.

5.7

#6 Xiaomi Releases MiMo-V2-Pro Model

Xiaomi has released its new model, MiMo-V2-Pro. While the model delivers decent performance, it is not considered to be at the technological frontier. Notably, the model does not utilize open weights, highlighting an emerging trend among Chinese AI developers who are increasingly moving away from open-weight releases for their advanced models.

5.6