Together AI

We just open sourced a set of agent skills for Together AI. The idea is simple: Instead of copy-pasting from docs every time you need to call an API, your coding agent already knows how. The skills cover chat completions, image/video/audio generation, fine-tuning, embeddings, batch inference, evaluations, dedicated endpoints, GPU clusters, and more. They work with Claude Code, Cursor, Codex, and Gemini CLI. One command to install: npx skills add togethercomputer/skills Each skill includes reference docs, runnable Python/TypeScript scripts, and routing logic so the agent picks the right one automatically.

New from Together Research: LLMs can fix the query plans your database optimizer gets wrong, resulting in up to 4.78x faster execution. Cost-based optimizers fail when they miss semantic correlations. A filter that prunes 15M rows to 2.9M gets applied after a join instead of before — because the optimizer assumed independence where none existed. DBPlanBench exposes DataFusion's physical operator graph to an LLM, which applies targeted JSON patch edits to fix join ordering without regenerating the full plan. On TPC-H and TPC-DS workloads: → Up to 4.78x speedup on complex multi-join queries → 60.8% of queries improved by more than 5% → Build memory: 3.3 GB → 411 MB on a single benchmark query → Plans optimized at small scale transfer directly to larger databases Paper and code are open-source. Link in comments.

We’re excited to announce Deepgram's speech and voice models on Together AI. AI natives can now deploy Deepgram’s STT and TTS models natively on Together AI Dedicated Model Inference and run the full real-time voice stack — from transcription to reasoning to synthesis — on the AI Native Cloud. Deepgram’s lineup includes Flux, Nova-3, Nova-3 Multilingual, and Aura-2, built for low-latency voice agents operating in real-world production environments. Highlights: → Conversational STT built for turn-taking — Flux delivers 250ms end-of-turn detection for faster, more natural exchanges → Production-ready transcription and synthesis — Nova-3, Nova-3 Multilingual, and Aura-2 support noisy audio, multilingual interactions, and clear enterprise voice output → Enterprise infrastructure by default — Dedicated workloads, 99.9% uptime SLA, zero data retention, SOC 2 Type II, HIPAA-ready support, and data residency options Read the announcement: https://lnkd.in/gVHwraYT

The Together AI kernels team pushes performance to the next level. An investigation into how left more questions than answers, but VP of Kernels Dan Fu seemed proud. If you want the full picture, read on: https://lnkd.in/gwtkFbU9

New from Together Research: Aurora Speculative decoding that adapts to shifting traffic in real time. Static draft models degrade as domains change. Offline retraining can't keep pace. Aurora fixes this — an open-source, RL-based framework that learns continuously from live inference traces without interrupting serving. Key results: → 1.25x over a well-trained static speculator through online adaptation → Online training from scratch surpasses a carefully pretrained baseline → Under abrupt domain shifts, Aurora recovers quickly You don't need extensive offline pretraining. Aurora learns from the first requests it serves. Code is open-sourced. Read the blog and paper [links in comments]

Open or closed — how do you choose the right AI model? For AI-native companies and enterprises, it's one of the most consequential decisions you'll make. Together CEO Vipul Ved Prakash takes the stage at HumanX with Mozilla's Mark Surman and The Wall Street Journal's Rolfe Winkler to work through what actually matters, because the model you choose now shapes what you build next. 🗓️ Thursday, April 9, 11:40 a.m.

New from Together Research: small models can beat GPT-4o on long context with the right system design. The instinct when context windows hit 128K or 1M tokens is to throw everything into one prompt. In practice, performance degrades as length grows. Our new paper, accepted at #ICLR2026, introduces a framework to study when and why "Divide & Conquer" works, and how to design it effectively. The core insight: long context failures come from three distinct noise sources: 1/ Model noise: confusion grows superlinearly with input length 2/ Task noise: chunks lose cross-document context 3/ Aggregator noise: the Manager fails to stitch partial answers correctly Naive "MapReduce" approaches collapse on that third point. The fix is a Planner agent that rewrites the task prompt so Workers return exactly what the Manager needs. Results: Llama-3-70B and Qwen-72B using this framework consistently outperform GPT-4o single-shot on retrieval, QA, and summarization as context length scales. The smaller models win, and they're cheaper and faster. The limit: tasks with high cross-chunk dependency — where a clue on page 1 connects to page 100 — still favor the single-shot approach. Blog, paper, and code in the comments.

We had a big week at #NVIDIAGTC 2026. Here’s what we shipped together with NVIDIA: • NVIDIA Dynamo 1.0 — Dynamo is already baked into our full-stack AI platform, and this release pushes inference performance even further for teams running production workloads.  • NVIDIA OpenShell via NemoClaw — We’re hosting the OpenShell runtime, so developers can tap into 150+ optimized models and build autonomous agents without sacrificing safety or scale.  • NVIDIA Nemotron 3 Super — This 120B parameter hybrid MoE model (just 12B active per token) brings long-horizon reasoning and multi-agent collaboration to production, and is now available via Together Dedicated Model Inference. • NVIDIA Parakeet TDT 0.6B V3 — Fast, reliable transcription meets Together’s inference infrastructure. Building real-time voice agents just got a lot more straightforward. Read it all here: https://lnkd.in/g-kxFKR8 Open innovation is what drives Together AI, and this is what it looks like in practice. #TogetherAI #NVIDIAGTC #OpenSource #AIInfrastructure #Inference

Together AI is heading to HumanX — and we're bringing the AI Native Cloud experience to booth #819. Stop by for live demos with our Solution Architects, customer story activations, research meet & greets, and a build-your-own hat bar and custom comic book station. We're also taking the stage twice: 📅 April 7 | Dan Fu, VP of Kernels — Building for compute that doesn't exist yet (but almost does) 📅 April 9 | Vipul Ved Prakash, CEO — Open or closed models: What AI natives & enterprises need to know Come find us, see what we've been building, and hear from real AI-native teams getting real results with Together AI. 👉 Booth #819. See you there: https://lnkd.in/gXZvTzky