Applied Compute

Applied Compute has released a 35B open-weight model, built on Qwen3.6-35B-A3B, that answers code search questions across roughly 9,000 repositories by querying a precomputed index. Developed in partnership with turbopuffer, the model is aimed at engineering teams and infrastructure providers who need repo-scale code retrieval without paying frontier API rates, and the company says it handles these queries at about 100x lower cost than a frontier model while topping the needle-in-a-haystack task outright. The launch matters because it puts a concrete number on the case for training smaller, task-specific models against your own index rather than routing every query to a general-purpose LLM. It also doubles as a proof point for Applied Compute's core pitch, which is that companies get better economics by owning post-training, evaluations, and data pipelines for narrow workloads rather than renting a general model. The startup trains custom AI models optimized for each customer's use cases, with every custom model trained on the data of the organization that commissioned it. Applied Compute was founded by Yash S Patil, Rhythm Garg and Linden Li, with Yash S Patil serving as CEO . Patil previously helped build agentic software systems including work tied to OpenAI's Codex, Garg was a core contributor on OpenAI's o1 reasoning model, and Li focused on machine learning infrastructure and reinforcement learning training systems. The turbopuffer collaboration is an early public example of how their platform, which the company positions around training, serving, and continuously improving customer-owned models, gets applied to a real retrieval workload.