A 35B open-weight model trained to search a precomputed index answers repo search questions at 100x lower cost than a frontier model. We partnered with @turbopuffer to train Qwen3.6-35B-A3B to find code across ~9,000 repositories. It tops the needle-in-a-haystack task outright at 2-10x lower latency.

“50% of DoorDash’s agentic restaurant orders are going to places users have never ordered from before.” @andyfang tells our CEO @ypatil125 what happens when agents become the discovery layer. If models increasingly decide what gets surfaced and bought, companies have a strong reason to train and own that intelligence.

@andyfang and our CEO @ypatil125 get into what actually counts as proprietary data. Sometimes it’s obvious, like customer behavior or merchant data. Other times it’s buried in a support agent hearing “happy birthday” and knowing to send a cake. The opportunity is to turn the judgment @DoorDash has accumulated over years of operating into proprietary intelligence it can own, train, and compound.

“Until you actually see things operationally, it’s going to be hard to build DoorDash from scratch.” Our CEO @ypatil125 sat down with @andyfang on why cheaper software doesn’t erase years of operating advantage. @DoorDash’s moat is its proprietary data, edge cases, and hard-won knowledge, and increasingly, the models trained on top of it.

At Kimi K3 scale, memory directly determines how many GPUs you need for training. For example, streaming gradients for Adam updates cuts the host memory peak by ~33%. Another bottleneck was SiTU-GLU activations. Naive autograd saves redundant tensors and consumes >100GB of HBM at ~100k tokens/GPU. A custom operator that streams through a fixed-size workspace saves ~90GB of HBM per GPU.

Over the last six months, AC2 has powered our Applied AI team’s work for Microsoft, NVIDIA, Cognition, Harvey, DoorDash, and others. Now, we’re excited to empower every AI team with their own model factory. AC2 is available in private beta. Reach out to learn more.

We partnered with @Harvey to train a model for Review Table, achieving state-of-the-art accuracy at a fraction of the cost and latency of frontier alternatives. Review Table is one of Harvey’s highest inference volume products. Training an open source model for production traffic required building a representative eval and understanding how real lawyers use the product. Our teams discuss the research process.

This represents a vital step towards closing the gap between training and inference, and moving towards continual learning in production. Several of our customers have been training with BYOH. Read the full report.

AC2 supplies a rollout-specific OpenAI-compatible endpoint backed by the policy being trained, plus a lifecycle wrapper to initiate, monitor, and grade rollouts. Our trace server reconstructs each rollout as a sequence of append-only episodes using structural prefix matching, so compaction and subagent calls stay grouped under one rollout with their distinct contexts preserved. Your harness keeps ownership of context, tools, and state; only completions and grading artifacts cross into the training stack.

You can now post-train a model inside your existing production harness with our platform, AC2. A production harness is a whole engineered system around the LLM, with its own context management, tools, sandboxing, and control flows. Porting that into a new training runtime can be expensive and could introduce train-test mismatch, where the policy is optimized against a simulated harness and then struggles in production. All you need to do is swap out the harness’ LLM response endpoint to one provided by AC2, and expose a lightweight protocol for AC2 to initiate and grade rollouts; the trainer handles the rest.

As a supporter of the open weights ecosystem, we're proud to be a post-training partner for NVIDIA Nemotron. We post-train Nemotron models for customer use cases, de-risk mainline RL runs on our AC2 platform and training stack, and contribute aggregate workload statistics for inference benchmarking. This is how open models get better, and we're excited to keep working closely with @NVIDIAAI.

Last year, reasoning models unlocked the ability to post-train agents for long horizon tasks. Being introduced to coding agents was quite a magical experience for me! I also thought that the unlock would lead to models being deployed for all kinds of other tasks. However, the broader deployment of AI seems to be hindered by the gap between the environments curated for RL training and the complex orchestration logic that is actually used in practice. Further progress requires creating increasingly complex synthetic environments and leveraging new techniques to improve agents based on their experiences in production. I joined Applied Compute to work on bridging that gap between theory and practice. Models can and should be useful for any task today; we just need to integrate the model with the infrastructure it needs for its tasks. We've translated our early training reps with customers into a platform built for agents that deliver outcomes. Really excited by the progress and the path forward to broader deployment. There's still a lot more work to be done!

Applied Compute is proud to co-sign the Open Weights and American AI Leadership letter. AI is a new form of IP. Every company should be able to own its intelligence, trained on its own data, shaped by its own experts. Open weights are what make that possible.

Open models got so good in 2025 that our vision of developing infrastructure to help customize models finally felt tractable. Before, it would have taken an prohibitively large amount of R&D and compute spend to pretrain from scratch and then develop a stack to serve our in-house models. Now we have a family of open weights models propped up by an ecosystem of hardware/software co-design that allows ourselves and the companies we work with the participate in AI development. We stand on the shoulders of a vibrant ecosystem of open models so we can focus on helping companies build better models enabled by their own data. Excited that @nvidia is leading the charge on this.

Frontier open models create the substrate for intelligence to be adapted, differentiated, and owned. We need them to preserve competitive markets across industries, where companies can imbue their models with the data and judgment that make them unique.

Applied compute has co-signed the Open Weights and American AI Leadership letter put forth by Microsoft and NVIDIA. AI is a new form of IP and it is important that every company and person is able to own their intelligence.

Your data is your edge, but only if your AI is built on it. Rent a generic model and so can your competitor. The companies with an edge are deploying custom models that they own and improve over time. Our co-founder @rhythmrg recently stopped by @southpkcommons to share how companies are owning their intelligence with Applied Compute.