June 2026 Summaries

Frontier AI research spans multiple layers, from agent development to GPU kernel optimization, ensuring improvements at each stage contribute effectively to the overall system, as demonstrated by the Together platform. Research efforts have produced notable advancements: the DSGym framework standardizes evaluation for data-science agents across diverse tasks, while ThunderAgent enhances agent inference efficiency and TTT-Discover achieves state-of-the-art discoveries using open models. Model shaping techniques like RARO and V1 enhance reasoning capabilities without relying on traditional verifiers, and algorithmic optimizations such as Aurora improve adaptive speculative decoding for dynamic traffic. Systems optimizations, including Untied Ulysses and OEA, address memory efficiency and MoE decode latency, and the ParallelKernelBench sets new benchmarks for multi-GPU kernel generation, emphasizing the importance of communication over computation. These innovations are presented at ICML 2026 in Seoul, where Together AI showcases its comprehensive, multi-layered research approach.

An experiment comparing the open-source Kimi K2.7 Code and the proprietary Claude Fable 5 for generating landing pages revealed that Kimi is significantly more cost-effective, being approximately 16 times cheaper, while still producing competitive quality. Both models initially produced AI-generated pages that lacked uniqueness, but by using a custom MCP server to provide design inspiration and visual references, the quality of Kimi's output improved markedly, with better hierarchy, typography, and composition. Despite Claude Fable scoring slightly higher on quality metrics, the cost-effectiveness of Kimi, combined with its capability to improve through visual context, demonstrated that open-source models are a practical option for creating landing pages. The results were documented and published on the OVSC website, highlighting the potential of open-source models when given strong inputs and the ability to iterate cost-effectively.

ParallelKernelBench (PKB) provides a benchmark and evaluation framework for multi-GPU kernel generation, focusing on replacing PyTorch + NCCL with CUDA kernels that utilize NVLink for data transfer. While large language models (LLMs) like GPT-5.5 and others have shown some proficiency in generating single-GPU kernels, their performance in multi-GPU scenarios is limited, with only a minority of generated solutions outperforming naive baselines. The evaluation highlighted key challenges such as reasoning about rank coordination, data partitioning, and optimal GPU-to-GPU transfer mechanisms, with most models relying on limited communication mechanisms and struggling with intricate hardware-specific abstractions. Although some models achieved notable speedups, particularly in familiar patterns like collective primitives, the overall findings underscore significant gaps in current LLM capabilities for multi-GPU kernel optimization. PKB's open benchmark aims to foster further research into AI-driven optimization and distributed infrastructure management, encouraging contributions and exploration into inter-node fabrics and other advanced architectures.

Together AI has achieved ISO 27001:2022 certification from A-LIGN Compliance and Security, Inc., confirming that its Information Security Management System (ISMS) meets the latest international standards for information security management. This certification underscores Together AI's commitment to providing secure and well-governed infrastructure for AI workloads, following a comprehensive assessment of its risk management, data security, and continuous improvement strategies. The certification encompasses the systems and processes protecting customer data and platform operations, covering both corporate headquarters and third-party data centers. ISO 27001:2022, recognized globally, assures customers of robust governance over sensitive information and secure practices, serving as a benchmark for evaluating Together AI as a critical infrastructure provider. The certification is part of a broader defense-in-depth security program that includes secure infrastructure, rigorous access management, and continuous monitoring, complementing existing controls aligned with other industry frameworks like SOC 2. Together AI continues to evolve its security programs to address emerging AI system risks, scaling controls and governance to accommodate more sensitive workloads.

Together AI has partnered with MiniMax to serve the newly launched MiniMax M3 model, a state-of-the-art model featuring advanced capabilities such as a 1M-token context window, native multimodality, and agentic workflow support. Together AI's Inference and Kernel teams have achieved significant engineering breakthroughs to efficiently serve M3 at scale, including the development of a KV-Block-Major sparse attention kernel, a novel paged attention integration for MSA, and a Rust-based multimodal preprocessing gateway. These optimizations have resulted in throughput improvements of 81-125% across different concurrency levels. The M3 model introduces MiniMax Sparse Attention (MSA) to overcome attention-computation bottlenecks, significantly boosting performance in long-context workloads. The collaboration between Together AI and MiniMax addresses engineering challenges such as supporting the 1M context length and optimizing multimodal processing. Additionally, Together AI will host the open-weights model, facilitating developer access upon its public release.