Together AI

It's been a very exciting month at Together AI! Catch up on all the updates to our product, research and tools in our latest "Together We Build" issue 📰

🏆 Together AI has been named to the 2025 Forbes AI 50 — Forbes' annual list of the most promising privately-held companies using AI to transform industries and shape the future. This recognition is a testament to our momentum and mission: to build The AI Acceleration Cloud — turbocharged by cutting-edge AI research and engineered for massive scale 🌐 From high-performance GPU clusters to simple developer APIs for inference and fine-tuning, Together AI powers advanced models like DeepSeek 🐋 and Llama 🦙 We’re proud to be recognized in #ForbesAI50 And we’re just getting started 🚀

ICYMI during NVIDIA GTC we announced Together Instant GPU Clusters ⚡ Up to 64 interconnected NVIDIA GPUs, available in minutes, entirely self-service, perfect for training models of up to ~7B parameters 🚂 , or running models like DeepSeek-R1 🐋 Now available in Preview, request access at 🔗 together.ai/instant

🧠💻 DeepCoder-14B – A fully open-source, RL-trained code model! 🚀 🔹 What is DeepCoder-14B? A 14-billion-parameter code generation and reasoning model jointly trained by Agentica and Together AI using Reinforcement Learning with verifiable rewards (RL-VR). It achieves 60.6% Pass@1 on LiveCodeBench, matching OpenAI’s o3-mini — with open source weights and training experiments. 🔹 Why does it matter? DeepCoder-14B sets a new standard for open models in code reasoning: ▪️ LiveCodeBench: 60.6% ▪️ HumanEval+: 92.6% Pass@1 ▪️ Codeforces Elo: 1936 (95.3 percentile) ▪️ All trained in just 2.5 weeks on 32×H100s. 🔹 How was it trained? Reinforcement learning with automated unit-test-based rewards 32K→64K context generalization using iterative context lengthening Training powered by the Together Code Interpreter, scaling to 1000+ test executions per minute 🔹 What is DeepCoder-14B good for? 🧠 Complex coding tasks ⚙️ Competitive programming 🧩 Long-context reasoning across structured problems 🔒 Enterprise-grade privacy & deployment DeepCoder-14B runs on Together AI’s ultra-fast inference engine with secure, private deployments, including options to run in your VPC and full control over data sharing. 📚 Read more: 📝Blog: together.ai/blog/deepcoder 🤗Model: https://lnkd.in/gECpSYq8 🖥️GithHub: https://lnkd.in/gWuN2FqH

🚀New on Together AI: Cogito v1 Preview on Dedicated Endpoints 🎉 We’re excited to announce support for the Cogito v1 Preview models — now available on Together Dedicated Endpoints! Developed by Deep Cogito, the Cogito v1 series includes powerful open-source LLMs up to 70B parameters, trained with Iterated Distillation and Amplification (IDA) for scalable alignment and enhanced reasoning. These models offer both direct responses and self-reflective outputs, making them well-suited for agentic and high-accuracy applications. By running Cogito on Together Dedicated Endpoints, you get: ⚡️ Single-tenant performance on your own cluster 🧠 Consistent low-latency inference at scale 🔧 Custom model & hardware configurations 💸 Cost-efficiency, with recent price drops of up to 43% Try it out and see what next-gen reasoning feels like → https://lnkd.in/gdPEsYJK

Open Deep Research – implemented from scratch in code! In this notebook, we share our Deep Research implementation & go over how it works step-by-step. We use an agent workflow capable of planning, search, and self-reflection. Notebook: https://lnkd.in/gDJ-4Jkh

You can now run Llama 4 directly on the Hugging Face model page – powered by Together AI!

🚀 Big news! We’re thrilled to announce the launch of Llama 4 models on Together AI as an AI at Meta launch partner. Both groundbreaking Llama 4 models, Llama 4 Maverick and Llama 4 Scout, are now available on the Together API. 1️⃣ Llama 4 Maverick has 17B active params and 400B total. It's a 128-expert MoE powerhouse for multilingual image/text understanding (12 languages), creative writing, and enterprise-scale applications—outperforming Llama 3.3 70B. It natively supports a 1M context window (500k currently supported on Together). 2️⃣ Llama 4 Scout is smaller at 17B active params and 109B total. It natively supports a 10M-token context (300k currently supported on Together), and excels at multi-document analysis, codebase reasoning, and personalized tasks. A smaller model than Maverick but state of the art in its size & with text + image input support. Get started with Llama 4 on Together AI today! 🔗 Read our announcement blog: https://lnkd.in/e6wXa9p2 🚀 Start using Llama 4 Maverick: https://lnkd.in/eeTKY3VD 🗺️ Try Llama 4 Scout: https://lnkd.in/eMaudv2B

🚀 Big news! We’re thrilled to announce the launch of Llama 4 models on Together AI as an AI at Meta launch partner. Both groundbreaking Llama 4 models, Llama 4 Maverick and Llama 4 Scout, are now available on the Together API. 1️⃣ Llama 4 Maverick has 17B active params and 400B total. It's a 128-expert MoE powerhouse for multilingual image/text understanding (12 languages), creative writing, and enterprise-scale applications—outperforming Llama 3.3 70B. It natively supports a 1M context window (500k currently supported on Together). 2️⃣ Llama 4 Scout is smaller at 17B active params and 109B total. It natively supports a 10M-token context (300k currently supported on Together), and excels at multi-document analysis, codebase reasoning, and personalized tasks. A smaller model than Maverick but state of the art in its size & with text + image input support. Get started with Llama 4 on Together AI today! 🔗 Read our announcement blog: https://lnkd.in/e6wXa9p2 🚀 Start using Llama 4 Maverick: https://lnkd.in/eeTKY3VD 🗺️ Try Llama 4 Scout: https://lnkd.in/eMaudv2B

New Notebook: Building AI Agents with Agno and Together! Create powerful agents that can search the web, access knowledge bases, and work together in specialized teams. Code and Docs below! Notebook: https://lnkd.in/guDf53PY Docs: https://lnkd.in/g3R9dAWV