HybridAgent-8B-Agentic

• Paper: Hybrid Reasoning Agents
• Code: https://github.com/togethercomputer/HybridAgent
• 14B HybridAgent: togethercomputer/HybridAgent-14B-Agentic

Overview

HybridAgent is an LLM agent that dynamically navigates between latent reasoning and text-based interaction within a single generation process. It thinks in continuous latent space through a lightweight HybridLink and acts through text, calling a Python interpreter and a web search tool across multiple turns over one shared working memory. Its latent thoughts stay directly interpretable: every latent step can be read out as pseudo-text.

This checkpoint is the 8B HybridAgent for agentic reasoning tasks, built on Qwen3-8B-Base with HybridSFT followed by HybridRL.

Files

Standard Transformers files for the backbone (model.safetensors, config.json, generation_config.json, the tokenizer with the two anchor tokens <latent> and </latent>) plus the HybridLink weights adapter.pt with adapter_config.json. Loading the backbone with transformers alone gives a text-only model; hybrid generation with latent mode and tools needs the HybridAgent code.

Usage

git clone https://github.com/togethercomputer/HybridAgent.git && cd HybridAgent
pip install -r requirements.txt && pip install -e .
cp .env.example .env   # TAVILY_API_KEY for the search tool

# batch inference on a JSONL file of questions
python -m hybrid_inference.run_batch --model togethercomputer/HybridAgent-8B-Agentic \
    --input examples/questions.jsonl --output outputs/results.jsonl

# inspect the latent thoughts of a trajectory as pseudo-text (HTML report)
python -m interpretable_latent_mode --model togethercomputer/HybridAgent-8B-Agentic \
    --question "How many trailing zeros does 100! have?" --output report.html

import torch
from hybrid_agent import HybridGenerationEngine, SYSTEM_PROMPT, build_tool_executor, load_hybrid_agent

model, tokenizer = load_hybrid_agent("togethercomputer/HybridAgent-8B-Agentic", device="cuda")
engine = HybridGenerationEngine(model, tokenizer, torch.device("cuda"), build_tool_executor(), SYSTEM_PROMPT,
                                "How many trailing zeros does 100! have?", batch_size=1, record_latent_readout=True)
engine.run(torch.Generator(device="cuda").manual_seed(0))
print(engine.row_result(engine.rows[0])["output_rendered"])

Citation

@article{zou2026hybrid,
  title   = {Hybrid Reasoning Agents},
  author  = {Zou, Jiaru and Qi, Yunzhe and Chen, Sirui and Bianchi, Federico and Zhu, Shang and
             Lu, Pan and He, Jingrui and Zou, James},
  journal = {arXiv preprint arXiv:XXXX.XXXXX},
  year    = {2026}
}