About Together AI

Company Overview

Together AI provides an AI-native cloud platform for open-source AI inference, fine-tuning, and deployment, co-founded by Vipul Ved Prakash with a $305M Series B.[bio] Named to Forbes AI 50 for its complete AI lifecycle support including fast inference and large-scale fine-tuning.[21] Topics: ai-infra, inference, open-source, gpu-cloud.[bio]

Model Deployments and Benchmarks

Together AI hosts production-ready open models with 99.9% SLA in serverless or dedicated modes.[1][2][9] DeepSeek V4 Pro delivers SOTA coding (93.5% LiveCodeBench, 3206 Codeforces, 80.6% SWE-Bench Verified) via hybrid attention (27% lower FLOPs, 10% reduced KV cache vs V3.2) and modes: Non-think, Think High, Think Max.[1][2][3] Challenges: Benchmarks self-reported without independent verification; SWE-Bench sensitive to agent frameworks; efficiency vs prior model only; SLA uptime-focused.[structured claims & counter-claims] Kimi K2.6 (Moonshot AI) offers multimodal agentic AI with 300-sub-agent swarms (80.2% SWE-Bench, 89.6% LiveCodeBench, 79.4% MMMU-Pro).[9][10][11] Other integrations: Alibaba Wan 2.7 for video gen/editing,[30][31][32][37] Deepgram STT/TTS for voice agents,[33][34][35][38] NVIDIA Nemotron 3 Super (hybrid MoE, 1M context).[46]

Research Contributions

At ICLR 2025, presents on model efficiency, long-context reasoning, next-gen attention, decoding.[4][5][6][7][8] Challenges: 'New work' may not confirm accepted papers; topics could be general focus.[counter-claims] Parcae enables stable looped transformers matching 1.3B quality at 770M params via spectral radius control (e.g., 370M: 20.00 Core vs Transformer's 17.46). Scales recurrence/data via power laws for FLOP-efficient inference.[14-20] Mamba-3 SSM optimized for inference (better prefill/decode vs Mamba-2, Transformers).[44] Aurora: Online RL for adaptive speculative decoding.[41] Divide & Conquer boosts long-context perf with small models outperforming GPT-4o.[42] DBPlanBench: LLMs patch database query plans for speedups.[28][29][36]

Infrastructure and Optimizations

AI-native cloud for GPU-intensive workloads, rapid research-to-production.[26] Kernel team optimizes LLM perf (FlashAttention heritage), e.g., FlashAttention-4 hits 1605 TFLOPs/s on Blackwell.[39][40][47] Fine-tuning supports 100B+ models, tool-calling, VLMs.[43] Partnerships: NVIDIA (Dynamo, Nemotron),[45][46] OpenClaw agents.[27] EinsteinArena: AI agents solve math problems (e.g., kissing number 11D: 593→604).[12][13]

Blogs and Tools

• AI-Native Cloud[26]
• Kernel Optimization[40]
• Aurora SpecDec[41]
• D&C Long-Context[42]
• Fine-Tuning Update[43]
• Mamba-3[44]
• NVIDIA GTC[45]
• Nemotron[46]
• FlashAttention-4[47]