Stanford and Together AI's HybridAgent: 8B agent beats GPT-5.6 Luna on GAIA

Stanford, UIUC and Together AI released HybridAgent, an 8B and 14B Qwen3-based agent that can reason in a latent mode without emitting tokens. The 8B model scores 43.7 on GAIA versus 37.9 for GPT-5.6 Luna, but trails on HLE at 14.4 versus 29.0. Weights, data and code are MIT-licensed.

• GAIA: HybridAgent 8B scores 43.7, 14B 52.4, GPT-5.6 Luna 37.9
• HLE: 14.4 for 8B and 16.0 for 14B versus 29.0 for Luna
• Average 10.4% gain, 35.5–54.7% fewer tokens and 1.8–2.6x faster inference
• Latent mode feeds hidden state back via a ~10M-parameter HybridLink adapter without decoding