Introducing Dria-Agent-α, the first agentic LLM trained for Pythonic function calling:. - 3B parameters, yet matches GPT-4o in BFCL.- Parallel multi-function calls at once.- Reasoning beyond JSON's limitations.- Trained with synthetic data generated by 2000 edge devices

Compilers for the AI Age: Rethinking Inference Engines, by Stefan Teleman of @driaforall .

Fast, Private AI for Everyone: The Orion FHE Framework, by Brandon Reagen (Asst. Prof.) of @NYUniversity.

Synthetic Data, Real Reasoning: Lessons from the Frontier of Mathematical AI, by @ShubhamToshniw6 of @nvidia .

Humans in the Loop: Building Trustworthy Agentic AI, by @qadri_sarmad of @LastMile .

DriaConnect drew more than 200 attendees to @betaworks NYC last week. Explore sessions on agentic AI, synthetic data, distributed inference, and private AI featuring researchers from NVIDIA, NYU, LastMile, and more below.

RT @LastMile: Packed house last night at @betaworks: 500+ signed up, 200+ showed up IRL 🙌. Our founder @qadri_sarmad shared updates on mcp-….

Reserve your spot by registering here:.

NYC July 1: . Join us to explore private/distributed inference & agent-native platforms. Talks:. Brandon Reagen (FHE Framework for Deep Learning), . @ShubhamToshniw6 (Synthetic Data), . @awrigh01 (AI Driven Venture Capital), . Stefan Teleman (Distributed Inference).

Blog: By leveraging the mathematical properties of transformers and a carefully designed three-party protocol, our goal is to provide:. - Real-world deployability with minimal overhead. - Provable privacy against realistic threat models. - Perfect.

Introducing our research on low-overhead, privacy-preserving distributed inference: . Protecting inputs, outputs, and model weights using lightweight permutations, enabling efficient and privacy-preserving LLM inference at scale.

Start using Batch API today: Blog:

Launching Decentralized Batch Inference—cutting costs by 5x+ vs. traditional providers—and hitting 20,000 daily active users contributing hardware globally. Dria Batch API is perfect for synthetic data, offline model eval, and cost-sensitive batch jobs.

Blog: Guided generation with our approach outperforms pure randomness, as diversity increases with additional iterations and nodes. More nodes enable broader, parallel exploration; however, saturation necessitates deeper inference to reach rare.

We introduce a decentralized, guided-generation pipeline using classifier-free guidance, dynamic negative prompts, and iterative regeneration to asynchronously scale diversity, overcoming duplicates and static curricula in LLM-based synthetic data.

Join us today at our event in the Bay Area to explore more about Dnet:.

Read the full story on our blog:. Music for the video composed by our CTO, @andthatto.

Most of the world's energy flows through heterogeneous hardware, yet today's AI software is designed for homogeneous and centralized setups. The concentration of high-performance AI into the hands of a few powerful entities poses profound risks to humanity. Introducing Dnet: a

Join us for a deep dive into open-source AI, RL, commodity hardware inference, and next-gen silicon. May 7th — we’re hosting:. - Kyle Corbit (@corbtt) from @openpipe, sharing insights on reinforcement learning and ART, an open-source RL library for agentic AI. - Peter

Want to join the Dria network but not sure where to start?. Check our new guide:. Or start with one-click node providers:.@easynodexyz .@mintairxyz .@xNodePro .@octa_space . Get your node up and running effortlessly—no coding required. Dria is just.

Dria’s Node Launcher has surpassed 500,000 downloads, with over 100,000 nodes launched alongside @ollama. Each download scales the network, bringing more programmable compute to the world.