A comparison of GLM-5.3 and its distilled variant, GLM-5.3 Flash, on the DeepSWE benchmark reveals that distillation preserves core coding capability while significantly reducing rollout costs. The full model achieves a 69.0% pass@1 rate at $3.99 per task, whereas the Flash variant scores 63.4% at just $0.24, representing a 17x price reduction.
Together AI has introduced Canary rollouts, a feature that migrates live traffic from a current model to a new checkpoint in gated steps. This system automates the safety mechanism for swapping models by running health checks and optional metric gates before moving traffic, allowing for automatic pauses and reversals if performance degrades.
Together provides a framework for enterprises to migrate from closed-source AI models to open-source models (OSM) using managed services, aiming to reduce complexity and accelerate adoption. The process involves discovering suitable models through benchmarks, evaluating them via traffic replay, adapting prompts or weights, and calculating ROI to justify the switch.
Together AI has expanded its fine-tuning service to support a wider range of open-weight models, including GLM-5.3 and Kimi K2.7, while introducing live experiment tracking and finer controls over the training process.
The kernels team at Together AI has optimized the ThunderKittens library to leverage new features in the NVIDIA Vera Rubin NVL72 platform, specifically targeting improved performance for NVFP4 and FP8 matrix multiplications.
Together AI announces the public preview of preemptible compute for Together GPU Clusters on Kubernetes, offering a lower-cost option for interruption-tolerant workloads. Preemptible nodes utilize unused NVIDIA accelerated capacity and are billed at a flat 50% of the on-demand rate.
This article outlines the MIGHT Stack (Model, Inference, Gateways, Harness, Tools), a modular architecture that allows developers to leverage open source models for agentic software without needing deep machine learning expertise or dedicated hardware.