Predibase

#AI use cases are narrow: a specific outcome needs to be achieved with your model. So why use oversized, slow general-purpose #LLMs that do an "okay" job on your task? 🤔 Teams that are serious about #production GenAI understand that small task-specific models (#SLMs) deliver the best possible performance for their use case with a much smaller, more #efficient footprint. 👣 We're honored to be recognized by Gartner as a GenAI #CoolVendor for leading the charge in small task-specific model training and serving. 😎 Check out Gartner's latest Cool Vendors report to learn more 👇 Thank you Arun Chandrasekaran, Manjunath Bhat, Arup Roy and George Brocklehurst for all of your great research! 👏

Lots of headlines floating around about #DeepSeek-R1, but the biggest takeaway is the dominance of #opensource 💪 There's arguably no reason to be locked into #commercial models anymore—with open-source you can own your models, run them #faster at a lower cost, and get better performance by #finetuning 💡 For more perspectives on what DeepSeek means, check out the latest blog from Greylock featuring the perspectives of founders from AI companies on the bleeding edge incl. Devvret Rishi (Predibase), Alexander Ratner (Snorkel), Tuhin Srivastava (Baseten), Jerry Liu (LlamaIndex), and Ankur Goyal (Braintrust). Check it out: https://lnkd.in/gh8Z2ZHU

Want to see how #DeepSeek distills perform out of the box? You can deploy DeepSeek-R1-Distill-Qwen-32 privately and securely with just a few clicks once you're in Predibase. Sign up for a free trial to get $25 of inference credit to give it a shot today! Free Trial: https://lnkd.in/gA9ZZksN Read more about #DeepSeek deployments: https://lnkd.in/gmk4Hpu7

⚡ Not all #inference is created equal ⚡ When #LLM traffic for customer-facing apps spikes, rigid #GPU deployments can't keep up or need to be over-provisioned. Is your inference stack providing smart #autoscaling? If not, you're missing out: 🐎 High throughput / low latency when demand surges 💰 30%+ cost savings when traffic dips Don’t get stuck with #idle GPUs. Scale smarter and serve your customers better. Download our Definitive #Guide to Serving Open-source LLMs for more best practices: https://lnkd.in/gv2UJQrw

🚀 Deploy #DeepSeek Models in Your Cloud or Ours in Seconds DeepSeek-R1 and its distilled variants are some of the most powerful open-source models available today, but many teams are struggling to deploy them efficiently in their own cloud (#VPC). It doesn’t have to be complicated. With Predibase, you can: ✅ Deploy DeepSeek-R1 or any of its distillations like DeepSeek-R1-Distill-Qwen-32B in your VPC with a few lines of code ✅ Spin up a private SaaS deployment in seconds using efficient L40S (enterprise teams can quickly access H100s and H200s) ✅ Ensure enterprise-grade security & compliance with dedicated infrastructure 🔥 Special Offer: The first 10 qualified organizations we speak to can fine-tune DeepSeek distills for #free and get 20% off a one-year deployment! 🔗 Read the full guide on deploying DeepSeek: https://lnkd.in/gzfJJeGJ #LLMs #DeepSeek #MLOps #AI #Inference #Predibase

🐋 Everyone’s talking about #DeepSeek-R1—and for good reason. It’s the first open-source model to go toe-to-toe (and sometimes win) against top commercial models. But let’s be real… The big question everyone’s asking us: Can you #finetune it? Short answer? Yes. Longer answer? It’s complicated. But we’re doing it today at Predibase. Reasoning models like DeepSeek-R1 are difficult to improve using traditional supervised fine-tuning (SFT), and require the use of new #reinforcement learning (RL) based techniques (like #GRPO) to improve performance on specific tasks. At Predibase, we've developed a new framework for efficient LoRA based reinforcement learning, combining the latest RL techniques with efficient and scalable infrastructure built on top of Predibase and LoRAX, allowing you to customize DeepSeek-R1 and its variations for your real-world data and tasks. We’ve spent the time experimenting (so you don’t have to), and we’re sharing everything in a webinar on Feb 12 @ 10AM PT. 🔍 What we’ll cover: ✅ How to fine-tune DeepSeek-R1-Qwen-7B with reinforcement learning ✅ Benchmarks on how RL fine-tuning impacts reasoning performance ✅ When to use a reasoning model vs. sticking with a generalist SLM If you’re trying to push your models further, this one’s for you. Register here → https://lnkd.in/gSicrGaQ See you there 🤝

🤔 Asking yourself why #DeepSeek-R1 is in every headline lately? It's not just about the performance gains (they're huge!) or about being 27x cheaper than #OpenAI's o1 reasoning model—DeepSeek-R1 is shaking things up with its novel approach to #reinforcement learning (RL). This isn't just another tweak; it’s a game changer. 💡 What’s the big deal with RL? It's about learning from interaction, especially when training data is scarce or we've already tapped out all available data. Think of it as learning by doing—less cramming from textbooks, more hands-on problem solving. Could this be the technique that takes over AI in 2025? 🔓 Open source and transparent, DeepSeek-R1 lets you see under the hood with reasoning traces for every response—important for anyone in industries where trust is everything. 🌟 Dive into our DeepSeek-R1 teardown post and see what the buzz is all about. https://lnkd.in/gYemt_qu #AI #DeepSeekR1 #ReinforcementLearning #FutureOfAI #TechTalk

Agents. Agents. Agents. ‼️ Sure they're cool, but what does it take to actually build #AI agents that work? Come find out 🔥 Last call to save your spot for our interactive session and live #demo with Siddharth Ghatti, Enterprise Architect at Marsh McLennan to get an inside look at what it takes to build production-grade agents! Check out what you'll learn: - How #finetuning for function calling dramatically improves intent recognition - Why fine-tuned #SLMs outperform commercial LLMs in agentic systems - How Multi-LoRA serving (#LoRAX) accelerates + economizes Agentic AI - Behind the scenes look a real production agentic workflow Save your spot: https://lnkd.in/gFW5z8qd

High accuracy is a must when building #agentic systems. Full stop. Why? Compounding Errors 🛑 In an agentic workflow, each agent executes a different step to accomplish the task. If your model’s #accuracy is 90% at each step, then your overall accuracy drops to 60% when compounded over just 5 steps. So how can you build more accurate #AI agents? With fine-tuned #SLMs! 🔥 Want to learn more? 👇 Join us on Wednesday to hear how #MarshMcLennan built a highly accurate AI Assistant with #finetuned agents and what it takes to get started on your own. 🎟️ Save your spot: https://pbase.ai/4jBtjGm 🎟️ Image credit: Ardent Venture Partners.

That's it. That's the social post. #DeepSeek Try it today: https://lnkd.in/g7jHUDVX