Research Vault

Together HybridAgent 14B Agentic

Properties1
tagshuggingface, text-generation, agentic, tool-use, reasoning, latent-reasoning, hybrid-reasoning, research, instruct

4 min read

Together HybridAgent 14B Agentic

One-line: Together AI’s research agent that interleaves latent reasoning steps with text and tool calls in one generation stream (Python interpreter + web search across turns), post-trained from Qwen3-14B-Base with HybridSFT then HybridRL — MIT, but it needs the HybridAgent runtime: stock transformers loads only a text-only (base) backbone.

Coding benchmarks

BenchmarkScoreSource
HumanEval—not published
HumanEval+—not published
MBPP—not published
LiveCodeBench—not published
Aider polyglot—not published
SWE-bench Verified—not published
SWE-bench Multilingual—not published

The paper evaluates math/science/medicine/search/agentic reasoning — there is no repo-editing or polyglot-edit lane at all, so the classic coding cells are genuinely absent, not merely unlisted.

Agentic / tool-use benchmarks

BenchmarkScoreSource
BFCL—not published
τ-Bench—not published
ToolACE—not published
GAIA (Lv.1 / Lv.2 / Lv.3 / Avg)53.8 / 55.8 / 33.3 / 52.4paper Table 1, lab-reported — vs Kimi-K3 41.7 avg, DeepSeek-V4.1-Flash 39.8, ARPO(14B) 43.7, GRPO(14B) 36.9, GLM-5.3-Flash 30.1, QwQ-32B 18.9, ReAct 32.0
WebWalker (Easy / Med / Hard / Avg)48.9 / 52.4 / 46.5 / 49.5paper — vs ARPO(14B) 40.5, GRPO(14B) 38.5
HLE (NS / CE / SF / Avg)16.1 / 13.3 / 17.9 / 16.0paper — vs ARPO(14B) 10.0, GRPO(14B) 8.6
SealQA (S-0 / Hard)24.3 / 37.8paper — vs ARPO(14B) 17.1 / 22.8
DeepSearchQA (Acc / F1)21.7 / 46.8paper — vs ARPO(14B) 16.3 / 38.2

Reasoning (context only)

BenchmarkScoreSource
MMLU / MMLU-Pro—not published
GPQA Diamond55.1paper Table 2 (RL stage) — vs GRPO 52.5, ARPO 54.0
MATHMATH500 88.2paper — vs GRPO 87.4, ARPO 87.0, standard text SFT 83.8
AIME 202543.3 (avg@16)paper — vs GRPO 40.8, ARPO 42.1
AIME 202645.8 (avg@16)paper — vs GRPO 42.5, ARPO 43.3
MedQA81.7paper — vs GRPO 81.0
MuSiQue34.5paper — vs GRPO 32.0, ARPO 31.5

Efficiency

  • Active params: 14B dense (no MoE); the only added weights are the HybridLink adapter.pt projector.
  • Context length: Qwen3-14B-Base native (32,768); the paper evaluates with a shared multi-turn working memory, not a long-context extension.
  • VRAM (fp16 rough): ~28 GB bf16 backbone + adapter — comfortably single-GPU, and the cheapest scale in the family.
  • Efficiency claims (lab, paper): 35.5–54.7% fewer text tokens than matched text agents, 1.8–2.6× faster inference, 2.7–3.7× lower cost; the cost table (Table 9) lists a 4B-scale comparison of 51.2 → 13.9 TFLOP per question.
  • Deployment caveat: the card is explicit — loading the backbone with plain transformers gives a text-only model, and the backbone is Qwen3-14B-Base (not instruct). Hybrid latent mode, the tool executor (Python + web search, Tavily key) and the interpretable latent read-out all require git clone https://github.com/togethercomputer/HybridAgent && pip install -e .. For a drop-in agent loop, plain Qwen3-14B-Instruct-class models are simpler; this one is for the hybrid-latent experiment.

What makes it notable

HybridAgent is the open release of a genuinely different agent training recipe rather than another SFT/DPO pass: the model switches between continuous latent thoughts and text within one autoregressive stream, keeps a lossless hybrid working memory across turns, and can read its latent steps back out as pseudo-text for debugging. Training is HybridSFT (learn to coordinate with the HybridLink projector) then HybridRL (learn when to think in latent space vs act in text), on Qwen3-8B/14B/4B-Base backbones. The paper reports the 14B checkpoint beating the strongest agentic baseline in its own suite by 17.4% on average and matched text/latent baselines by 10.4% at the RL stage, at a fraction of the tokens. Limits to weigh: the GAIA Lv.2 figure (55.8) is higher than the frontier closed baselines the paper itself measures (GPT-5.6-Luna 32.7, Kimi-K3 42.3) — a lab-harness result that no third party has reproduced, and Artificial Analysis has no entry; there is no SWE-bench/Aider/LiveCodeBench/BFCL number at all, so this is not a repo-editing agent; and the useful mode needs the custom runtime. It qualifies here on the “novel post-training algorithm + published agentic evals” rule, not on drop-in usability.

See also

Graph View

Backlinks