Tracking Our Latest Research, Systems Work, And Real-World Performance Wins
Interactive tool-agent post-training is unstable because user simulation noise corrupts delayed-reward learning. We propose EigenData + verifier-based RL, achieving strong improvements on τ²-bench (Airline 58.0% → 73.0%; Telecom 53.7% → 98.3%).
Eigen AI and Boson AI are co-hosting the Higgs Audio Hackathon at Circuit Launch in Mountain View, March 20–22, 2026. Build with Higgs Audio and Eigen AI's full model suite.
Introducing Mini Banana, a new approach to AI.
We are proud to announce that we have expanded this performance leadership across 11 of the most prominent models in the industry, consistently securing top-tier rankings for speed, latency, and efficiency.
Palo Alto, Calif., Jan 15, 2026 - The future of voice has arrived: Eigen AI and Boson AI join forces to power Higgs-Audio v2.5.
Eigen AI announces full open-source support for DFlash — covering both training and inference. DFlash uses block diffusion for up to 6.17× lossless acceleration.
Production-grade EAGLE-3 draft model checkpoints (SpecBundle Phase 1) plus SpecForge v0.2 upgrades for training and scaling speculative decoding.
Introducing VISTA-Gym and VISTA-R1 for tool-integrated visual reasoning, achieving up to 18.72% gains across 11 benchmarks.
High Quality, High Throughput, Low Latency, Strong speed-for-price.
Fine-tuning Qwen-Image-Edit on Pico-Banana-400K unlocks four-step Lightning edits with EigenTrain, EigenInference, and EigenDeploy.