August 2025 Summaries

DeepSeek-V3.1, now available on Together AI, is a hybrid model that offers both fast response and deep reasoning modes, accessible through configurable chat templates. Users can select non-thinking mode for quick responses or thinking mode for detailed analysis, with the model providing comparable quality to DeepSeek-R1 but with enhanced speed, making deep reasoning feasible for production. It includes built-in support for code and search agents, optimized for multi-turn tool-calling workflows, and is backed by extensive training for robust performance in extended conversations and large codebases. The model excels in various applications, from generating API endpoints and fixing syntax errors to debugging complex systems and designing multi-step research workflows. Together AI ensures reliable performance with its serverless APIs and infrastructure optimized for large mixture-of-experts models, offering tools for easy integration, such as a Python SDK, to deploy this model in production environments.

Building AI agents to automate complex and long-running engineering tasks demands a distinctive approach compared to typical AI applications, as evidenced by a case study on using agents to enhance LLM inference through speculative decoding. The blog outlines key patterns for developing effective agents, emphasizing the importance of infrastructure and behavioral patterns. Infrastructure patterns focus on creating a stable environment with well-abstracted tools and comprehensive documentation, while behavioral patterns guide agents on managing tasks like parallel sessions, wait times, and progress monitoring. A real-world example illustrates the automation of a speculator training pipeline, highlighting the reduction of manual oversight and increased efficiency. Despite challenges such as context management, novel failure modes, and resource optimization, the potential for expanding AI agent usage into various domains remains vast, promising increased reliability and enhanced human-LLM collaboration.

OpenAI's release of the gpt-oss-120B and gpt-oss-20B models, licensed under Apache 2.0, represents a significant milestone in AI development by offering fully open-weight language models designed for customization. Together AI facilitates the fine-tuning of these models, allowing organizations to create AI systems tailored to specific domains, workflows, and requirements without the complexities of managing distributed training infrastructure. The Together AI platform simplifies the fine-tuning process into three steps—uploading datasets, configuring training parameters, and launching jobs—while automatically handling technical challenges such as data validation and memory allocation. The platform supports high-performance deployment with enterprise-grade reliability, including SOC 2 compliance and a 99.9% uptime SLA. Fine-tuning these models not only improves performance and cost efficiency for specialized tasks but also ensures stability and control over the application's lifecycle, free from external dependencies.

Parsed, in collaboration with Together AI, demonstrates how small open-source models can outperform large proprietary models on complex tasks, such as healthcare scribing, through task-specific optimization and comprehensive evaluation. By employing a rigorous evaluation-first methodology and fine-tuning models for specific tasks, Parsed achieves 60% better accuracy with 10 to 100 times lower inference costs compared to larger models. This approach allows for greater transparency and reliability, essential in domains like healthcare where precision is critical. Parsed's advanced evaluation harnesses align closely with expert judgment to ensure clinical soundness, source fidelity, and adherence to specific styles, enabling smaller models to exceed the performance of larger, general-purpose models. The partnership with Together AI provides a robust fine-tuning platform that supports continuous optimization and seamless deployment, allowing organizations to achieve superior performance and significant cost savings.

OpenAI's release of the gpt-oss series marks a significant step in open-source AI, featuring two models: a 20B parameter model and a 120B model, which is purportedly comparable to proprietary alternatives. Together AI has integrated these models into their platform, highlighting the benefits of open-source AI, such as complete model ownership, cost-effectiveness, and deployment flexibility without vendor lock-in. The gpt-oss-120B model was tested against the o4-mini model across five practical scenarios, including code generation, creative tasks, instruction following, mathematical reasoning, and information synthesis. The results favored gpt-oss-120B, especially in complex instruction-following and practical coding tasks, demonstrating its competitive edge and strong reasoning abilities. This open-source model provides developers with the freedom to fine-tune and deploy AI models at a fraction of the cost of proprietary ones, positioning itself as a viable alternative for a wide range of applications in the evolving AI landscape.

OpenAI has released two new open-weight models, gpt-oss-120B and gpt-oss-20B, under Apache 2.0 licensing, which allow for unlimited modification and deployment across various industries and regions. Available through Together AI, these models are designed for reasoning tasks and offer significant performance and economic advantages. Together AI provides a robust infrastructure with 99.9% uptime, comprehensive developer tooling, and seamless scalability, making these models suitable for enterprises, emerging markets, government applications, research, and education. The models are purpose-built to be highly customizable, enabling organizations to deploy them in diverse environments while maintaining high safety standards. With significant collaboration from NVIDIA to enhance performance, Together AI ensures immediate access to these models, offering a platform that is both reliable and cost-effective for deploying AI applications.