Together AI

Huge thanks to the attendees, guest speakers, and sponsors who joined #AINativeConf at such a defining moment. As builders, researchers, and founders ship new AI-Native products and shape what’s possible – we’re excited for the future of AI and its applications. Thank you for coming along with us.

Congrats to the @cursor_ai team on Composer 2 — a huge milestone for RL-trained models and step forward for open-source coding intelligence. Together AI is proud to partner on this launch. Composer 2 is turning heads for its speed and quality — and we help power the Composer 2 Fast endpoint on the AI Native Cloud. Excited to keep building with the Cursor team.

Day 3 of #NVIDIAGTC: trivia champions crowned, lightning talks from XY.AI + Hypertec, and a deep dive on AI-native apps at scale. Catch up on everything we announced this week 📷 together.ai/blog/together-…

Read more: together.ai/blog/fine-tuni…

What's new: 👉 Tool call fine-tuning with end-to-end OpenAI-compatible schema validation 👉 Reasoning fine-tuning with native thinking token support 👉 Vision-language model fine-tuning for domain-specific visual data 👉 Up to 6x throughput gains on MoE models with cost and time estimation before and during training

Together Fine-tuning now supports tool calling, reasoning, and vision-language model fine-tuning. Train models up to 1T parameters with up to 6x higher throughput on MoE architectures.

.@togethercompute is bringing its inference research with NVIDIA Dynamo 1.0 to deliver an accelerated, cost-effective inference stack for production AI workloads.

Day 2 of #NVIDIAGTC brought the heat — literally 📷 Hot wings, a lightning talk from 5C, Tokens After Hours with @Metronome Webhook, and our team met Jensen. Not a bad Tuesday. Day 3 kicks off soon — Together Trivia, cool prizes, Booth #1213. Come ready to booth #1213. 📷📷

Paper: arxiv.org/abs/2603.15569 Code: github.com/state-spaces/m… Blog: together.ai/blog/mamba-3

Introducing Mamba-3 🐍 Inference speeds are more important than ever, driven by the rise in agents and inference-heavy RL rollouts. Linear models are fast in FLOPs but memory-bound during decode. Mamba-3's MIMO (multi-input, multi-output) variant fixes this: swap the recurrence from vector outer-product to matrix multiply, and you get a stronger model at the same decode speed. Fastest prefill+decode at 1.5B. Beats Mamba-2, GDN, and Llama-3.2-1B. Kernels open-sourced. #mamba3 #togetherresearch Congratulations to the team leading this research: @aakash_lahoti @kevinyli_ @_berlinchen @caitWW9 @tri_dao @_albertgu

Day 1 of #NVIDIAGTC wrapped. Talks, demos, hats, comics — and the best community we could ask for. 📷 Back tomorrow with more at booth #1213.

We’re less than 3 days away from #NVIDIAGTC! Join us for sessions featuring leaders across Together AI, @DecagonAI, and @cursor_ai where we dive into how purpose-built cloud infrastructure accelerates AI innovation — and how Together powers the next generation of AI-native applications.

Here are more resources on how we built open deep research & the code! - Blog post explaining it: together.ai/blog/open-deep… - GitHub repo with the code: github.com/Nutlope/open-d…

Try out the Open Deep Research v2 app here! opendeepresearch.dev

Introducing v2 of our Open Deep Research app! Generate detailed reports on any topic with open source LLMs. Fully free & open source. We're releasing everything: evaluation dataset, code, app, and blog 🔥

The world’s leading AI infrastructure platforms are converging on the same voice model 🔥 Excited to announce that Cartesia is now a dedicated model partner on @togethercompute's Voice Platform for the 450K+ teams and developers building on Together.

Real-time voice agents are getting fast enough to feel conversational🎙️ MiniMax Speech 2.6 Turbo is now part of the voice stack on @togethercompute

Most voice stacks today are stitched together across vendors. Together puts the whole pipeline in one place for natural, real-time conversation. Here is how it works: together.ai/blog/build-rea…

What that means for builders: 👉 Co-located STT, LLM, and TTS on one cloud so every handoff stays inside one cluster 👉 @Cartesia and @DeepgramAI models now hosted natively on Together infrastructure 👉 Swap models across the full stack without rebuilding integrations 👉 One unified surface for billing, deployment, and access

Today, Together AI is launching a unified solution for building real-time voice agents with the entire pipeline running on one cloud. AI natives can now deploy voice apps for every use case at production scale.