Brief

Together AI has launched Together Instant Clusters, a new service providing readily available, self-service GPU clusters for AI development and deployment. This offering aims to simplify the complex process of setting up multi-node GPU infrastructure, allowing users to provision clusters with hundreds of GPUs in minutes via API, CLI, or console. The service includes pre-configured components for distributed training and inference, supporting NVIDIA's latest GPU architectures and high-performance networking solutions. AI

Together AI has significantly upgraded its Batch Inference API, introducing a more user-friendly interface and expanding model compatibility to include all serverless and private deployment models. The update dramatically increases rate limits by 3000x, from 10 million to 30 billion enqueued tokens per model per user, enabling much larger-scale data processing. These enhancements aim to make high-throughput workloads more cost-effective and accessible, with costs typically at 50% of their real-time API for most serverless models. AI

MiniMax AI has released over 600 new voices through its Speech 2.8 Turbo model. These voices are now accessible on the Together AI platform. This expansion aims to provide a wider range of synthetic speech options. AI

Together AI has launched Violin, an open-source video translation tool designed to make online video content accessible across language barriers. The system utilizes advanced AI, including speech recognition, large language models, and speech synthesis, to provide high-quality translations. Violin also features interactive capabilities like a content-aware chat assistant and personalized voice selection, aiming to broaden the reach of video content globally. AI

Together AI has partnered with Cursor AI to launch Composer 2.5, a significant advancement for agentic coding models. This new version is noted for its speed and quality, pushing the boundaries of what coding agents can achieve. AI

Together AI has partnered with Pearl Research Labs to integrate blockchain technology into AI inference costs. This collaboration introduces a new inference endpoint for the Gemma-4-31B-it-pearl model, offering a discount of over 25% by offsetting costs with cryptocurrency emissions from the Pearl Network. The Pearl Network uses a Proof of Useful Work mechanism, where GPU computations for AI tasks simultaneously generate a cryptocurrency called PRL, aiming to fundamentally alter the economics of AI by reducing the price per token. AI

Together AI has released Gemma-4-31B-it-Pearl, an open-source model with enhanced capabilities. This model supports a 32K context window, configurable thinking processes, function calling, and JSON mode. It marks Together AI's initial offering powered by Pearl, with plans to broaden their Pearl-based product line in the future. AI

Together AI has released Gemma-4-31B-it-Pearl, an instruction-tuned model based on Gemma 4 31B. This model integrates the Pearl Network's Proof of Useful Work protocol, which generates proofs from existing matrix multiplications during training and inference. Users can access this model via a serverless inference endpoint on Together AI, with a discount on costs. AI

Together AI's speech-to-text models have achieved the top two positions on the Artificial Analysis leaderboard for transcription speed. The NVIDIA Parakeet TDT 0.6B V3 model, running on Together AI, is currently ranked first, processing 303 seconds of audio for every second of computation. AI

Together AI has launched Voice Finder, a new tool designed to help developers quickly select the most suitable voice for their applications from a catalog of over 600 options. The tool allows users to search for voices by describing their desired characteristics or by uploading an audio sample for comparison. Voice Finder categorizes each voice across more than 15 attributes, including pitch, accent, and emotion, to streamline the selection process for voice agents. AI

Together AI has detailed the architectural innovations behind DeepSeek-V4's ability to handle a 1 million token context window. The model employs a hybrid attention design that compresses context before storing it in the KV cache, significantly reducing memory pressure. This architectural shift transforms the challenge of long-context inference from a model capability into an inference systems problem, requiring optimized serving engines to manage cache layouts and batching effectively. AI

Together AI has launched a new feature allowing developers to deploy and run any model from Hugging Face using their Dedicated Container Inference (DCI) infrastructure. This is facilitated by an agent-based CLI tool called Goose, which automates the complex setup process, including inference server configuration and container generation. The system aims to significantly reduce the lag time between a model's release and its practical use, as demonstrated by the rapid deployment of Netflix's void-model. AI

Together AI has partnered with Adaption, a company co-founded by former Cohere and Google DeepMind leaders Sara Hooker and Sudip Roy. This collaboration integrates Adaption's data optimization tools with Together AI's fine-tuning infrastructure. The partnership aims to streamline the process for users to create high-quality, fine-tuned open-source models by improving dataset quality and simplifying the experimentation and deployment workflow. AI

Together AI has detailed its rapid response to a critical Linux kernel vulnerability, dubbed Copy Fail (CVE-2026-31431), which allows unprivileged local users to gain root access. The company treated the issue as a fleet-level emergency, disabling the vulnerable crypto socket interface across its infrastructure within hours. They also implemented a temporary kernel hardening step by unloading the vulnerable module and removing it from the module path, preventing its re-activation until stable upstream patches could be rolled out and tested. AI

Together AI has introduced Parcae, a novel stable architecture for looped language models. This new design allows models to achieve the quality of larger Transformers while using significantly fewer parameters, by increasing recurrence rather than solely scaling data. Parcae demonstrates improved stability over previous looped models and establishes the first scaling laws for this type of architecture, suggesting a more efficient frontier for training memory-constrained on-device models. AI

Together AI has launched the Wan 2.7 model suite, offering advanced video generation and editing capabilities. This suite includes text-to-video generation and will soon expand to image-to-video, reference-to-video, and video editing functionalities. The models provide users with greater creative control through features like audio-driven generation, frame-level conditioning, and reference inputs, all accessible via a unified API on the Together AI platform. AI

The Together AI kernels team, including researchers Dan Fu and Tri Dao, developed FlashAttention, a software layer that significantly optimizes GPU performance for AI models. This breakthrough, achieved by applying database system principles to GPU memory movement, resulted in 2-3x speedups, challenging the notion that transformer attention was already fully optimized. The team's subsequent work, including the ThunderKittens library, aims to accelerate kernel development for new hardware like NVIDIA's Blackwell GPUs, addressing the critical software-hardware gap in AI infrastructure. AI

Researchers at Together AI have developed a "Divide and Conquer" framework that enables smaller language models to effectively handle long context tasks. Their study, presented at ICLR 2026, demonstrates that by breaking down large inputs into smaller chunks and assigning them to multiple, less powerful models, performance can match or even surpass that of a single, large model like GPT-4o. This approach mitigates issues like model confusion and task-specific noise, leading to more efficient and cost-effective processing of extensive documents or codebases. AI

Together AI has enhanced its fine-tuning service to better support advanced AI workflows. The update includes native support for tool call, reasoning, and vision-language model fine-tuning, addressing common issues like unreliable tool execution and degraded reasoning in complex interactions. These improvements aim to increase iteration speed and accuracy for AI teams building agentic applications, with enhanced throughput and larger dataset handling for models up to 1T parameters. AI

Together AI has enhanced its GPU clusters with new features aimed at improving efficiency and manageability for AI-native teams. The platform now supports multi-tenancy, allowing different teams to share compute resources securely and independently. Key additions include autoscaling for elastic capacity, robust observability tools, and self-healing capabilities to reduce downtime and operational overhead. AI

Researchers at Together AI have found that current state-of-the-art speech recognition models exhibit a significant failure rate, averaging 39% error in transcribing street names, particularly for non-native English speakers who are 18% more likely to be misunderstood. This inaccuracy can lead to substantial real-world consequences, such as increased travel time and costs for services like ride-sharing. The study suggests that a synthetic data generation technique called "cross-lingual style transfer" can improve transcription accuracy by up to 60% with minimal training data. AI

Together AI has launched two new Rime models, V3 Turbo and V3, designed for natural code-switching in voice agents. V3 Turbo offers English-Spanish switching with a time-to-first-audio of approximately 120ms on dedicated endpoints, maintaining conversational flow and prosody. The V3 model supports switching across 11 languages, providing a unified solution for multilingual customer interactions without the need for separate language-specific models. AI

Researchers have introduced DSGym, a new framework designed to standardize the evaluation and training of data science agents. This system addresses limitations in current benchmarks by providing a unified API and self-contained execution environments, ensuring fair comparisons and enabling agents to utilize underlying data. DSGym integrates existing benchmarks and introduces new datasets for bioinformatics and machine learning competitions, demonstrating its utility by training a 4B parameter model to state-of-the-art performance among open-source agents. AI

Cursor, an AI-powered coding platform, has partnered with Together AI to optimize its real-time inference capabilities. This collaboration focuses on achieving low-latency responses within the editor's feedback loop, which is crucial for the AI's predictive and refactoring features. The partnership leverages NVIDIA's Blackwell architecture, specifically the GB200 NVL72, to enhance performance and reduce response times for developers. AI

Training large foundation models necessitates distributing the workload across numerous GPUs housed in multiple interconnected machines, a process known as multi-node training. This approach is essential for handling models with billions or trillions of parameters that exceed the memory capacity of single servers and would otherwise take months to train. Effective multi-node training relies on sophisticated parallelism strategies, high-speed network interconnects, and robust fault tolerance mechanisms to ensure efficient computation and progress. AI

Together AI has integrated Rime's enterprise-grade voice models, Arcana v2 and Mist v2, into its platform. Arcana v2 offers expressive, conversational voices trained on real customer interactions, while Mist v2 provides deterministic pronunciation control for high-volume applications. These models are designed to improve the reliability and naturalness of AI-powered voice agents, reducing latency and enhancing customer trust by ensuring consistent pronunciation and a more human-like conversational flow. AI

Together AI has released the Release Candidate for its new Python SDK, version 2.0. This updated SDK is built with a modern, type-safe architecture using OpenAPI specifications and Stainless, aiming for improved performance and easier maintenance. It replaces the legacy v1 SDK and introduces new features like Instant Clusters beta APIs, while also offering better type safety and editor support for developers. AI

Together AI is enhancing its cloud platform to support advanced reinforcement learning (RL) pipelines, integrating TorchForge and Monarch for distributed training. The platform now offers low-latency GPU communication and heterogeneous scheduling for mixed CPU/GPU workloads, crucial for complex RL tasks. New integrations with Together CodeSandbox and Code Interpreter allow RL agents to interact with tools and execute code, expanding their capabilities beyond traditional game-playing scenarios. AI

Researchers at Together AI have developed AutoJudge, a novel method to accelerate large language model inference. This technique automates the curation of task-specific datasets, enabling lossy speculative decoding without manual annotation. AutoJudge identifies critical tokens that impact downstream quality, achieving up to a 2x speedup over standard speculative decoding with minimal accuracy loss. AI

Together AI has launched a new startup accelerator program specifically designed for companies building AI-native applications. The accelerator will provide selected startups with platform credits, engineering expertise, go-to-market support, and access to a venture capital network. This initiative aims to help these startups scale their AI-native apps effectively on Together AI's platform. Early participants include Corridor.dev and PlayerZero, who are already leveraging the program's resources. AI

Together AI has appointed Mahadev Konar as its new SVP of Infrastructure Engineering to bolster its GPU cloud services. Konar, a key figure in Apache Hadoop's development and formerly VP of Infrastructure at Instacart, will lead efforts to enhance the reliability, performance, and scalability of Together AI's platform. The company aims to provide AI-native startups with a robust infrastructure, enabling them to focus on product development rather than managing complex GPU environments. AI