June 2025 Summaries

This guide provides a practical approach for developers to build their own coding agents from scratch using large language models (LLMs), function calling, retrieval-augmented generation (RAG), code execution and LLM workflows. The goal is to enable developers with the conceptual foundations and code-level insights needed to create their own agent pipelines. By combining tool use, context-aware retrieval, and runtime execution, these agents can go beyond text generation to act as real assistants in software development. The guide covers the key components of a coding agent, including function calling, code retrieval using embedding models, code execution with a safe sandboxed environment, and choosing the right workflow architecture. A real-world example is provided through a data science agent that demonstrates these capabilities in practice.

Together AI has announced a strategic alliance with Hypertec and 5C Group to deliver large-scale infrastructure for AI innovation across Europe. This partnership marks Together AI's largest infrastructure deployment in Europe, with initial rollouts planned for late 2025 and large-scale buildouts planned through 2028. The expansion combines infrastructure deployment with local team building, allowing Together AI to serve European customers' growing demand for sovereign, regulation-ready AI infrastructure. The partnership will deliver up to 2 gigawatts of AI-dedicated data center capacity, nearly 100,000 NVIDIA GPUs, and priority deployments in France, the UK, Italy, and Portugal. This deployment addresses compute capacity needs while enabling innovation across the region, with a focus on sustainability and regulatory compliance. The European deployment brings Together AI's full platform capabilities, optimized for local regulatory and operational requirements, including model ecosystem support, regulatory compliance, and sustainability focus.

The blog post outlines the process of building an effective data scientist agent using Together's open-source models and Together Code Interpreter (TCI). The agent is designed to handle multi-step data science tasks by leveraging the ReAct framework, which combines reasoning and action, allowing the agent to generate Python code snippets for execution. The implementation is modular, enabling easy modifications through prompt engineering while ensuring safe code execution with TCI's sandboxed environment. The agent's performance is evaluated using benchmarks like MLE-bench and DABStep, demonstrating competitive results, particularly in solving straightforward tasks. The post emphasizes the importance of robust execution environments, iterative design, and comprehensive testing in developing reliable AI agents. Despite limitations such as limited control over the agent's actions and minimal logging, the implementation serves as an accessible guide for building reasoning-driven AI assistants with open-source tools.

The Together AI Batch API offers businesses and developers a cost-effective solution for processing large volumes of LLM requests efficiently. By using batch processing, users can process non-urgent workloads at half the cost of real-time inference, with most batches completing within 24 hours. The API supports up to 50,000 requests in a single batch file, has simple integration with JSONL files, and provides real-time progress tracking. With the Batch API, users can scale their AI inference without scaling their budget, and pricing is based on successful completions at an introductory 50% discount.

The development of AI models has led to a seismic shift in platform technology, with open-source AI becoming increasingly ubiquitous and desirable. This shift is driven by the flexibility and cost-effectiveness of open-source models, which offer greater adaptability and lower costs compared to proprietary alternatives. The frontier for AI is rapidly expanding, with open-source models claiming numerous breakthroughs and holding multiple superlatives in the ecosystem. However, proprietary models are still advancing in areas such as reasoning and multimodality, but the gaps between them and open-source models are shrinking. To harness the power of open-source AI, a new kind of cloud platform is needed, which Together AI aims to provide by normalizing variability, providing adaptable abstractions, and adopting the latest AI-native infrastructure. As the demand for open-source AI continues to grow, companies like Salesforce, Zoom, and SK Telecom are already joining forces with Together AI, and developers are taking note of its potential.

YAQA (Yet Another Quantization Algorithm) is a new weight-only LLM post-training quantization method that quantizes models to directly preserve the original model's outputs. YAQA achieves state-of-the-art performance on downstream tasks by reducing the KL divergence to the original model by over 30% compared to existing rounding algorithms. It uses a near-optimal Kronecker-factored approximation of each linear layer's Hessian with respect to the KL, which is then used to quantize models with theoretical guarantees. YAQA has been shown to outperform existing methods in experiments, including reducing the cost of training by 20%, increasing network compression by 117x, and achieving faster training times by 4x.