Together AI News | Sep 06, 2024
Together AI announced two algorithmic improvements, including MagicDec, to speed up LLM inference...
Brought to you by RivalSense - an AI tool for monitoring any company.
RivalSense tracks the most important product launches, fundraising news, partnerships, hiring activities, pricing changes, tech news, vendors, corporate filings, media mentions, and other developments of companies you're following 💡
Company name: Together AI
Company website: together.ai
Company Overview
Together AI is a research-driven artificial intelligence company that contributes leading open-source research, models, and datasets to advance the frontier of AI. Their decentralized cloud services empower developers and researchers to train, fine-tune, and deploy generative AI models.
Latest Updates
- Together AI announced two algorithmic improvements, including MagicDec, to speed up LLM inference.
- 01.AI has made Yi-Coder, an open-source AI coding assistant, available through Hugging Face and a web interface at llamacoder.together.ai.
- Together AI's new TEAL method achieves 40-50% activation sparsity in Llama-2/3 and Mistral models, significantly enhancing decoding speeds.
- Together AI integrated the Chinese LM benchmark CLEVA into the HELM framework in collaboration with Liwei Wang's team from CUHK.
- Together AI released OLMoE, a 100% open-source Mixture-of-Experts LLM with 1B active and 7B total parameters, matching more costly models like Gemma and Llama.
- Together AI has added Qwen 2 models to their offerings, including Qwen 2 (1.5B), Qwen 2 (7B), and Qwen 2 (72B).
- Retool now supports custom AI providers, including Together AI.
- Together AI's new RAG reranker, LlamaRank, achieves a 92.9% hit rate and offers improved performance and interpretability.
- Together AI now offers the state-of-the-art LlamaRank reranking model exclusively on their API for RAG pipelines and other ranking tasks.
- Together AI launched a Rerank API compatible with Cohere and exclusive access to Salesforce's LlamaRank model, outperforming Cohere Rerank v3 and Mistral-7B QLM.
- Together AI announced Together Rerank + Salesforce LlamaRank with 8K context length, handling semi-structured data, improving search relevance, and reducing LLM compute costs in RAG systems.
- Together AI's cloud platform was integrated into the CAMEL-AI project as part of their latest sprint.
- Together AI is integrating with the distribution to set up as a remote inference provider for llama-stack.
- Together AI announced the beta release of Together Analytics, allowing users to view usage metrics like requests, latency, and TPM in their dashboard.
- Together AI will soon offer the Jamba 1.5 models from AI21 for deployment.
- Together AI's Llama Coder, powered by Llama 3.1, allows instant code publishing and sharing for quick prototyping or learning to code.
- Together AI extended LLaMA-2 to 32K context using Meta's Position Interpolation and their own data recipe and system optimizations.
- Together AI launched a project that helps generate and run code directly on the page.
- Together AI's integration is featured in the new Vectara-agentic package for building Agentic RAG applications.
- Together AI developed a robust validation framework to ensure hardware quality before deployment for AI startups and Fortune 500 companies.
- Together AI integrated Helicone into their inference service to enhance AI observability and transparency.
- Together AI has upgraded its state-of-the-art training clusters to include Nvidia H200 GPUs.
- Together AI has gated their optimized SwiGLU to their own use, moving away from open sourcing.
- Together AI offers up to 75% cost savings with 99.9% uptime SLA and flexible commitment options for their GPU Cluster.
- Together AI shared their latest work on speculative decoding for high-throughput inference on September 5.
- Together AI collaborated with SCB10X to develop ThaiExam, the first public leaderboard for language models on Thai scenarios.
- Together AI published new research titled 'TEAL: Training-Free Activation Sparsity in Large Language Models' on August 28, 2024.
- Ryan Pollock has joined Together AI as Director of Product Marketing in the United States, previously working as Director of Product Marketing in San Francisco.
- Together AI phased out the Mixtral 8x22B base model on 2024-08-28.
- Together AI announced plans to add new, more advanced models to their serverless platform.
- Together AI deprecated their best base models with 1 week's notice due to low usage and availability of newer models.
- Together AI partnered with Daily to bring open models to Daily Bots, enabling developers to build voice-to-voice AI and real-time multimodal agents with Llama 3.1.
- Hassan El Mghari from Together AI launched NSFW Filter, an npm library to block NSFW images, which has blocked over 100k images in the past year.
- Vipul Prakash, CEO of Together AI, partnered with Sahara AI which just raised $43 million.
- Together AI developed a robust validation framework for GPU clusters to ensure reliability for AI/ML workloads.
- Together AI will be speaking at the SCBX Next Tech event at Siam Paragon, sponsored by Amazon Web Services.
- Together AI partnered with Applied Digital to ensure powerful, secure, and dependable GPU clusters through a rigorous acceptance testing process.
- Shuaiwen Leon Song, VP of Research at Together AI, will be giving a keynote talk on 'Together Inference' at ACL 24.
- Tri Dao of Together AI will be a keynote speaker at the PyTorch Conference 2024 in San Francisco on September 18-19.
- Together AI will support the AI/ALL Summit: Future of Equitable and Inclusive AI at ACL 2024.
- Together AI will be hosting a recruitment mixer at The Black Cat restaurant on August 12 during the ACL meeting.
Did you find it useful?
If you liked this report, consider following your own companies of interest. Receive weekly insights directly to your inbox using RivalSense AI.