Back to Model
H

HybridAgent

Model Oct 19, 2026

Open Qwen3-based research agents that alternate continuous latent computation with text tool calls and include a nearest-token trace inspector.

Visit primary link 3 sources attached
HybridAgent authors' diagram shows the residual HybridLink adapter and alternating text and latent computation steps.
Author-published architecture illustration, with the full frame preserved. Fast and Inspectable is the authors' characterization, not an independently verified speed or transparency guarantee. Source: HybridAgent authors / togethercomputer, MIT
agent-ecosystemself-hosted

What happened lately

HybridAgent timeline

HybridAgent combines Qwen3-Base backbones with a learned HybridLink adapter and a custom inference engine. It performs intermediate steps in continuous latent space, switches to text for tool calls and answers, and carries mixed context across turns. Public Agentic checkpoints are available at 8B and 14B sizes under an MIT license.

The Python 3.11 setup uses the project’s source installation and adapter files. Loading the backbone with ordinary Transformers provides text-only inference rather than the full hybrid runtime. Search-enabled execution needs a Tavily API key, and the documented tool path also includes Python execution. An included inspector projects each latent step onto nearby token embeddings and produces a terminal view or standalone HTML trace. These readouts aid inspection; they are not a complete semantic decoding or a general safety guarantee. The original research paper was submitted October 13, 2026.