December 2025 Summaries

Together AI introduces MiniMax Speech 2.6 Turbo, a cutting-edge text-to-speech model designed to unify voice quality and low latency across multiple languages on a single platform, eliminating the need for disparate providers. This model is built on the infrastructure of Together AI, providing the capability to run alongside LLM and STT workloads with sub-250ms latency, ensuring real-time conversational interactions. MiniMax Speech 2.6 Turbo excels in naturalness, ranking highly on TTS leaderboards and offering automatic emotional awareness and multilingual support, accommodating over 40 languages with seamless inline switching. The platform caters to various applications such as customer service, content generation, and interactive entertainment, providing a cohesive infrastructure that supports a wide range of use cases while offering features like 10-second voice cloning and SOC 2 Type II and HIPAA compliance. This integration ensures reliable, native-quality speech synthesis with unified metrics and observability, making it an ideal choice for global applications requiring consistent voice quality.

Together AI's platform now integrates Rime Arcana v2 and Mist v2 into its Model Library, enhancing text-to-speech (TTS) capabilities with expressive, conversational voices suitable for customer service. Arcana v2 offers over 40 voices in multiple languages, trained on real customer interactions to improve conversational flow and provide measurable business outcomes, such as increased sales and reduced call abandonment. Mist v2 provides deterministic pronunciation control, essential for high-volume environments, ensuring consistent pronunciation across millions of calls. Both models operate on Together AI's unified infrastructure, which facilitates seamless integration with large language models (LLM) and speech-to-text (STT) processes, maintaining production-grade latency and compliance standards. This setup allows global contact centers, healthcare, and financial institutions to deploy advanced TTS solutions with consistent performance and regulatory compliance, leveraging Together AI's dedicated endpoints and infrastructure.

In the ongoing debate about the limits of digital computation in advancing artificial general intelligence (AGI), Dan Fu presents an optimistic view in his analysis titled "Yes, AGI Can Happen – A Computational Perspective." While acknowledging existing constraints like hardware limitations and stalled GPU progress, he argues that current AI systems have not yet reached their theoretical potential. Fu highlights the underutilization of hardware, noting that state-of-the-art training often achieves only about 20% Mean FLOP Utilization, with inference utilization even lower. He suggests that significant efficiency gains can be realized through improved software-hardware co-design and innovations such as FP4 training. Additionally, Fu points out that current AI models are based on outdated hardware, and the impact of new, larger clusters of GPUs is still forthcoming. Despite this, existing models are already significantly transforming complex processes, indicating that even without major technological leaps, there is substantial room for growth and improvement in AI capabilities.

NVIDIA's Nemotron 3 Nano, available on Together AI, is a cutting-edge solution for agentic and multi-agent systems, delivering high-quality reasoning at production speed. Utilizing a hybrid Mamba–Transformer and sparse Mixture-of-Experts architecture, it efficiently handles long-range dependencies and structured tasks while activating only a fraction of its parameters per token to improve speed and cost. With a 1M-token context, it supports complex, reasoning-intensive applications such as long-horizon planning and persistent agent memory. Together AI enhances Nemotron 3 Nano's performance with reliable, low-latency inference and scalability across agentic workloads, ensuring cost efficiency and flexibility with simple APIs. This combination empowers developers to create specialized agentic AI systems with transparency and scalability, making it ideal for diverse applications like coding assistants, scientific reasoning agents, and enterprise assistants.

Together has released the Python SDK v2.0 Release Candidate, a type-safe, OpenAPI-driven client aimed at enhancing speed, maintainability, and future-proofing for upcoming developments. This release invites developers to test the SDK during its beta phase to ensure compatibility and identify any regressions, with ongoing fixes and enhancements expected, including potential breaking changes. The new SDK boasts a modern architecture generated from OpenAPI specifications, offering better type safety and development experience akin to TypeScript, and introduces a host of new features including a revamped error model, modern parameter and type system, Jobs API for job management, Hardware API for discovering hardware options, and raw response and streaming helpers for improved debugging and control. The migration guide provides a detailed overview of the changes, highlighting easy and medium-effort migrations, new capabilities, and key breaking changes, while the RC period focuses on gathering feedback to refine the release before the general availability launch.

Reinforcement learning (RL) is increasingly integral to training modern agentic models, enabling them to handle long-term tasks, interact with dynamic environments, and execute complex workflows involving reasoning and tool use. In response to these evolving demands, Together AI and the PyTorch team at Meta have partnered to create an open-source RL framework on the PyTorch platform, aiming to enhance the capabilities of agentic systems. This collaboration integrates advanced RL into the Together AI platform, offering a comprehensive suite of services, including supervised fine-tuning and high-performance inference, to facilitate the development and deployment of next-generation AI models. The partnership emphasizes open-source development to empower the AI community and invites collaboration to advance training techniques and agentic system capabilities.

The AI Native Cloud is advancing reinforcement learning (RL) systems by providing a flexible and scalable infrastructure that supports modern RL pipelines, which require more than just simple training loops. Utilizing the full PyTorch stack, including TorchForge and Monarch, it offers distributed training capabilities on Together Instant Clusters, optimized for low-latency GPU communication and consistent cluster setup. These clusters accommodate heterogeneous RL workloads by efficiently managing GPU and CPU resources, and support complex RL frameworks that combine GPU-bound computations with CPU-bound tasks. Together AI also integrates tools such as CodeSandbox for microVM environments and Code Interpreter for isolated Python execution, facilitating tool-use, coding tasks, and simulations. A demonstration showcases a TorchForge RL pipeline operating on these clusters, training a model to play Blackjack, highlighting the adaptability of the system to different models and tasks. This setup paves the way for a flexible, open RL framework in the PyTorch ecosystem, aiming to deliver high-performance RL services on the Together AI Cloud, with ongoing collaborations and developments in partnership with Meta.

Speculative decoding is an advanced method that accelerates token generation by using a small draft model alongside a larger target model, where the draft suggests potential next tokens and the target verifies them. AutoJudge, an enhancement of this method, introduces an automated system for identifying and accepting "unimportant" mismatches, which are differences that do not affect the final output's correctness. This approach eliminates the need for human labeling by using a small classifier trained on existing embeddings to predict the importance of mismatches. AutoJudge shows notable improvements in inference speed across various testing scenarios, such as mathematical reasoning and programming tasks, by allowing more tokens to be accepted per cycle with minimal accuracy loss. It integrates seamlessly with existing speculative decoding frameworks and demonstrates substantial throughput gains, particularly in bandwidth-limited scenarios. However, the speedup benefits are dependent on the specific task and the frequency of unimportant mismatches, suggesting that threshold tuning for the classifier may be necessary for optimal results.

The team has focused on enhancing their inference platform to become the fastest for running top open-source models, achieving significant performance improvements as validated by benchmarks from Artificial Analysis. The platform now ranks first in output speed for several demanding models, including GPT-OSS-20B, GPT-OSS-120B, and Qwen-3-235B-Instruct, delivering up to 2.75 times faster performance than competitors. These improvements result from an integrated approach, including advanced GPU hardware exploitation, kernel optimization, low-bit quantization, and a scalable speculative decoding algorithm. The team re-engineered the entire system architecture to maximize the potential of NVIDIA Blackwell GPUs, leveraging high-performance kernels and quantization strategies like FP8 and FP4 to maintain model accuracy while boosting speed. The use of a scalable draft-model training pipeline further supports this performance leap, allowing for efficient speculative decoders and high-acceptance draft models. The company remains committed to advancing open-source AI model performance and scalability, with ongoing research into faster generation strategies and hybrid quantization approaches.