Hybrid Reasoning Agents — LLM agents that think in latent space and act through text
Stanford/Together AI/UIUC/Meta preprint plus MIT-licensed code: a HybridLink module toggles the agent between silent latent reasoning and text mode (for tool calls and answers), with HybridSFT/HybridRL post-training.
Why it matters
First working architecture that unifies latent-space reasoning with native tool use while keeping latent thoughts renderable as inspectable pseudo-text; claims 10.4% performance gain, 35.5-54.7% fewer tokens, 1.8-2.6x faster inference.
What you could build with it
A research team or startup could train small agentic models that do their internal reasoning in latent space and surface text only for tool calls, cutting the token cost of long-running search or research agents by more than half.
Does it hold up?
Code, training scripts, and trained checkpoints are all public with quick-start docs, so researchers can run it today — but the headline 2.7-3.7x cost-saving claims are author-reported against undisclosed frontier baselines and await independent replication.
Built with Hybrid Reasoning Agents
- HybridAgent toggles latent reasoning with tools, cuts cost 3.7xaiweekly · Third-party writeup of the paper's claims; notes ranges are published without per-benchmark figures or named baselines.
- Hybrid Reasoning Agents: they think silently and write only for toolsalextech · Independent coverage; flags that the performance figures are author-stated and need verification.
- togethercomputer/HybridAgentgithub · 71 ★ · Official MIT code: hybrid generation engine, HybridSFT/HybridRL training, batch inference, interpretable-latent inspector; released HybridAgent-8B/14B-Agentic checkpoints on Hugging Face.
Learn more
First spotted on github: source.
More AI models
OpenAI publishes 722 AI-generated math manuscripts from an unreleased modelOpenAI released 722 mathematical manuscripts in 372 result families to GitHub under Apache-2.0, produced by an…models · JEV 0.74StepFun Step-5-Preview — open BF16 weights releasedStepFun published the full BF16 weights for its flagship Step-5-Preview agentic model (600B-parameter sparse MoE, 27B…models · JEV 0.73ElevenLabs Eleven v4 + v4 Turbo: new emotive TTS architecture with a 100ms real-time variantElevenLabs released Eleven v4 and Eleven v4 Turbo on September 28: a new text-to-speech architecture with inline…models · JEV 0.72Mistral Large 4 (Le Chonk): 1.05T-param open-weight MoEMistral's new flagship: a 1.05T-parameter mixture-of-experts multimodal model with 49B active params and 1M-token…models · JEV 0.72VoiceStudio: open-source, fully-local ElevenLabs alternative with voice cloning, dubbing and dictation in 646 languagesVoiceStudio is an open-source, fully-local voice AI studio: voice cloning from a 3-second sample, voice design, video…models · JEV 0.71JetBrains Mellum 2.1: 12B MoE coding model trained for agentsJetBrains released Mellum 2.1, a 12B MoE model (2.5B active parameters, 128K context) under Apache 2.0, trained mainly…models · JEV 0.69
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.