Together AI

🚀 Together AI Sets a New Bar: Fastest Inference for DeepSeek-R1-0528 We’ve upgraded the Together Inference Engine to run on NVIDIA Blackwell GPUs—and the results speak for themselves: 📈 Highest known serverless throughput: 334 tokens/sec 🏃➡️ Fastest time to first answer token: 7.3 sec ⏱️ Lowest end-to-end response time: 9 sec Need more performance? Our Dedicated Endpoints hit 386 tokens/sec. Contact us to customize an NVIDIA HGX B200 deployment optimized for speed, quality, and cost. Our in-house stack delivers best-in-class performance and throughput—full stop. Agentic AI. Advanced reasoning. Blazing speed. Together AI is deploying Blackwell GPUs to power the next generation of real-world AI. (Read more - link in comments)

Introducing FutureBench — a new benchmark for evaluating AI agents on their ability to forecast future events that we developed with Hugging Face Rather than pattern-matching known answers, agents must synthesize data, reason under uncertainty, and make verifiable predictions about the future—like: 🔮 Will the Fed cut interest rates by 0.25% by July 1? 🗳️ Will Zohran Mamdani win by >13% in the NYC primary? 🌍 Will Ukraine and Russia hold peace talks next month? By drawing on news events and Polymarket, FutureBench generates dynamic, meaningful tasks that stress-test true agentic reasoning. Initial Results: 🥇 Agent Claude 3.7 Sonnet: 67.3% 🥈 Agent GPT-4.1: 62.0% 🥉 Agent DeepSeek-V3: 61.8% → See how Claude, GPT-4, and DeepSeek-V3 tackle the future: Explore the live leaderboard (link in comments) This is how we move beyond benchmarks—and get closer to building AI’s that reason, not just remember. #AIagents #FutureBench #LLM #Benchmarking #OpenSourceAI

🔥 Zain Hasan teamed up with Andrew Ng to teach a RAG course on Coursera! Together AI partnered with DeepLearning.AI to create this comprehensive 5-hour course covering information retrieval and search, LLMs, evals, and production scaling. All assignments, labs and technical demos are powered by open source models on the Together platform - giving developers hands-on experience with the same infrastructure powering production RAG systems. This is what happens when AI infrastructure meets world-class education 🚀

Here’s a peek behind the curtain at the early phase of our 36,000 liquid-cooled NVIDIA Blackwell GB200 NVL72 GPU cluster deployment, built on Dell Technologies integrated rack scale systems. Co-built with Together AI, this cluster delivers next-generation infrastructure to accelerate model training, fine-tuning, and inference at scale without compromising on performance, cost, or reliability.    As we continue to scale up the deployment, we’ll be sharing more updates along the way!    Bringing cutting-edge AI infrastructure to life takes careful planning and a ton of hands-on effort from our dedicated teams and technology partners—a big thank you to Hypertec Group, VAST Data, and CaTECH Systems Ltd.

Discover how to take retrieval-augmented generation (RAG) and search to the next level with Mixedbread’s latest open-source release: mxbai-rerank-v2. In this session from the "Learning Together" Series, Mixedbread CEO Aamir will present key insights into these next-generation reranking models, which are purpose built for high performance across diverse retrieval tasks. What you’ll learn: • How mxbai-rerank-v2 achieves benchmark-leading results using reinforcement learning • Practical steps for integrating these models into your own RAG or search systems • Performance insights across multilingual text, code, and tool retrieval scenarios • How improved retrieval quality simplifies enterprise AI workflows Register to receive the recording!

🚨 𝗞𝗶𝗺𝗶 𝗞𝟮 𝗶𝘀 𝗟𝗜𝗩𝗘 𝗼𝗻 𝗧𝗼𝗴𝗲𝘁𝗵𝗲𝗿 𝗔𝗜 𝗳𝗼𝗿 𝗽𝗿𝗼𝗱𝘂𝗰𝘁𝗶𝗼𝗻 𝘂𝘀𝗲 🚀 This 1T parameter powerhouse from Moonshot AI is outperforming proprietary models while delivering the cost control and transparency that enterprises actually need. Why this matters: ✍️ 𝗕𝗲𝘀𝘁-𝗶𝗻-𝗰𝗹𝗮𝘀𝘀 𝗰𝗿𝗲𝗮𝘁𝗶𝘃𝗲 𝘄𝗿𝗶𝘁𝗶𝗻𝗴 – #1 on EQ-Bench3 and Creative Writing benchmarks 🤖 𝗕𝘂𝗶𝗹𝘁 𝗳𝗼𝗿 𝗽𝗿𝗼𝗱𝘂𝗰𝘁𝗶𝗼𝗻 𝗮𝗴𝗲𝗻𝘁𝘀 – Native tool use, autonomous workflows, CLI integration 🏆 𝗦𝗢𝗧𝗔 𝗽𝗲𝗿𝗳𝗼𝗿𝗺𝗮𝗻𝗰𝗲 across LiveCodeBench v6, AIME 2025, MMLU-Redux, SWE-bench 💰 𝟲𝟬-𝟳𝟬% 𝗰𝗼𝘀𝘁 𝘀𝗮𝘃𝗶𝗻𝗴𝘀 vs comparable proprietary models Available now on Together AI: ⚡ 𝟵𝟵.𝟵% 𝘂𝗽𝘁𝗶𝗺𝗲 with leading price/performance optimization 🌍 Secure 𝗡𝗼𝗿𝘁𝗵 𝗔𝗺𝗲𝗿𝗶𝗰𝗮𝗻 𝗶𝗻𝗳𝗿𝗮𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲 with 𝗦𝗢𝗖 𝟮 𝗰𝗼𝗺𝗽𝗹𝗶𝗮𝗻𝗰𝗲 🛝 𝗧𝗿𝘆 𝗶𝘁 𝗶𝗻𝘀𝘁𝗮𝗻𝘁𝗹𝘆 in our playground ⚗️ 𝗙𝗶𝗻𝗲-𝘁𝘂𝗻𝗲 with Kimi Base for custom use cases 💰 𝗕𝗮𝘁𝗰𝗵 𝗔𝗣𝗜 for cost-effective distillation and synthetic data This is what the future looks like: frontier-level AI that's open, controllable, and economically sustainable. With ❤️ and 🙏 from Together AI to the Moonshot AI team.

🚨 𝗞𝗶𝗺𝗶 𝗞𝟮 𝗶𝘀 𝗟𝗜𝗩𝗘 𝗼𝗻 𝗧𝗼𝗴𝗲𝘁𝗵𝗲𝗿 𝗔𝗜 𝗳𝗼𝗿 𝗽𝗿𝗼𝗱𝘂𝗰𝘁𝗶𝗼𝗻 𝘂𝘀𝗲 🚀 This 1T parameter powerhouse from Moonshot AI is outperforming proprietary models while delivering the cost control and transparency that enterprises actually need. Why this matters: ✍️ 𝗕𝗲𝘀𝘁-𝗶𝗻-𝗰𝗹𝗮𝘀𝘀 𝗰𝗿𝗲𝗮𝘁𝗶𝘃𝗲 𝘄𝗿𝗶𝘁𝗶𝗻𝗴 – #1 on EQ-Bench3 and Creative Writing benchmarks 🤖 𝗕𝘂𝗶𝗹𝘁 𝗳𝗼𝗿 𝗽𝗿𝗼𝗱𝘂𝗰𝘁𝗶𝗼𝗻 𝗮𝗴𝗲𝗻𝘁𝘀 – Native tool use, autonomous workflows, CLI integration 🏆 𝗦𝗢𝗧𝗔 𝗽𝗲𝗿𝗳𝗼𝗿𝗺𝗮𝗻𝗰𝗲 across LiveCodeBench v6, AIME 2025, MMLU-Redux, SWE-bench 💰 𝟲𝟬-𝟳𝟬% 𝗰𝗼𝘀𝘁 𝘀𝗮𝘃𝗶𝗻𝗴𝘀 vs comparable proprietary models Available now on Together AI: ⚡ 𝟵𝟵.𝟵% 𝘂𝗽𝘁𝗶𝗺𝗲 with leading price/performance optimization 🌍 Secure 𝗡𝗼𝗿𝘁𝗵 𝗔𝗺𝗲𝗿𝗶𝗰𝗮𝗻 𝗶𝗻𝗳𝗿𝗮𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲 with 𝗦𝗢𝗖 𝟮 𝗰𝗼𝗺𝗽𝗹𝗶𝗮𝗻𝗰𝗲 🛝 𝗧𝗿𝘆 𝗶𝘁 𝗶𝗻𝘀𝘁𝗮𝗻𝘁𝗹𝘆 in our playground ⚗️ 𝗙𝗶𝗻𝗲-𝘁𝘂𝗻𝗲 with Kimi Base for custom use cases 💰 𝗕𝗮𝘁𝗰𝗵 𝗔𝗣𝗜 for cost-effective distillation and synthetic data This is what the future looks like: frontier-level AI that's open, controllable, and economically sustainable. With ❤️ and 🙏 from Together AI to the Moonshot AI team.

We just launched a new "dictate" feature on Together Chat powered by our new Whisper model! The video is not sped up – it's really that fast!

🚀 Live from Paris at the RAISE Summit, Together AI’s CEO Vipul Ved Prakash joined NYSE Wired & theCUBE to share how his team is redefining the future of AI infrastructure. Together AI, founded just three years ago, is building an AI acceleration cloud delivering supercomputing power for large-scale model training and inference, optimized down to the GPU. With tens of thousands of NVIDIA Blackwell GPUs being added every quarter. Their platform offers 200+ open-source models ready for fine-tuning, empowering developers with the flexibility and price-performance they crave. As AI-native companies and enterprises rush to unlock new use cases: from generative media to robotics and healthcare Together AI’s decentralized, regulation-ready infrastructure is paving the way for sovereign AI and the next wave of digital transformation. Watch Vipul Ved Prakash's Full Interview Below: https://lnkd.in/gg3q-W4g Full Wired & theCUBE - RAISE Summit Coverage Below: https://lnkd.in/eVbk92D4 SiliconANGLE & theCUBE - John Furrier - Kwiri Yang - Quantum AI - Henri Delahaye - David Vellante - Kevin Hawkins - Laura Diorio - Rajan Sheth #AI #TogetherAI #Innovation #DataDriven #Inference #AI #Nvidia #GPU #Founderfriendly #NYSEWired #theCUBE #RaiseSummit #theCUBE #Innovation #TechLeadership #AIInfrastructure #Supercomputing #SovereignAI

🚀 We just launched speech-to-text APIs designed for real-time applications. Our Whisper V3 Large deployment delivers transcription 15x faster than OpenAI while maintaining full accuracy. When transcription happens in seconds rather than minutes, entirely new applications become possible. What this enables: 🎤 Real-time customer support call analysis 📊 Meeting insights delivered before participants leave the room 🤖 Voice agents that respond naturally instead of asking users to wait 🏥 Medical scribes that keep pace with doctor-patient conversations Key capabilities: ⚡ Handle files exceeding 1GB (vs OpenAI's 25MB limit) 🔧 Process 30+ minute calls seamlessly 💰 $0.015 per audio minute with 50+ language support 💻 Available now through our standard APIs with the same authentication and billing you're already using.