Research Vault

Together HybridAgent 8B Agentic

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

3 min read

Together HybridAgent 8B Agentic

One-line: The 8B rung of Together AI’s latent+text hybrid agent (Python interpreter + web search across turns), post-trained from Qwen3-8B-Base with HybridSFT → HybridRL under MIT — its headline claim is GAIA 43.7 avg, ahead of DeepSeek-V4.1-Flash (39.8) and Kimi-K3 (41.7) in the lab’s own harness; needs the HybridAgent runtime, not plain transformers.

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

No repo-editing or polyglot-edit lane is evaluated in the paper — the coding cells are genuinely absent.

Agentic / tool-use benchmarks

BenchmarkScoreSource
BFCL—not published
τ-Bench—not published
ToolACE—not published
GAIA (Lv.1 / Lv.2 / Lv.3 / Avg)51.3 / 44.2 / 25.0 / 43.7paper Table 1, lab-reported — vs Kimi-K3 41.7 avg, DeepSeek-V4.1-Flash 39.8, ARPO(8B) 38.8, GRPO(8B) 32.0, ReAct 32.0
WebWalker (Easy / Med / Hard / Avg)48.9 / 45.2 / 35.2 / 42.5paper — vs ARPO(8B) 32.0, GRPO(8B) 30.0
HLE (NS / CE / SF / Avg)14.2 / 8.0 / 20.0 / 14.4paper — vs ARPO(8B) 8.8, GRPO(8B) 7.8
SealQA (S-0 / Hard)22.5 / 35.8paper — vs ARPO(8B) 18.9 / 26.0
DeepSearchQA (Acc / F1)20.0 / 40.7paper — vs ARPO(8B) 18.0 / 37.2

Reasoning (context only)

BenchmarkScoreSource
MMLU / MMLU-Pro—not published
GPQA Diamond48.0paper Table 2 (RL stage) — vs ARPO 46.5, GRPO 43.9
MATHMATH500 84.6paper — vs ARPO 82.8, GRPO 82.0, standard text SFT 78.2
AIME 202539.2 (avg@16)paper — vs ARPO 37.1, GRPO 36.3
AIME 202643.3 (avg@16)paper — vs ARPO 41.5, GRPO 39.4
MedQA75.3paper — vs ARPO 74.7, GRPO 74.3
MuSiQue32.0paper — vs ARPO 28.5, GRPO 27.5

Efficiency

  • Active params: 8B dense (no MoE); only the HybridLink adapter.pt projector is added.
  • Context length: Qwen3-8B-Base native (32,768); multi-turn hybrid working memory instead of a long-context extension.
  • VRAM (fp16 rough): ~16 GB bf16 — the smallest checkpoint in this release family, laptop/workstation-friendly.
  • Efficiency claims (lab, paper): 35.5–54.7% fewer text tokens, 1.8–2.6× faster inference, 2.7–3.7× lower cost than matched text agents (cost table uses the 4B scale as its worked example).
  • Deployment caveat: plain transformers loads the backbone as a text-only base model (Qwen3-8B-Base, not instruct); latent mode, the Python/web-search tool executor and latent read-out require the togethercomputer/HybridAgent package. Ablation in the paper shows the reward term r_tool matters most here — removing it costs 11.2% on GAIA and 3.5% on WebWalker.

What makes it notable

Same recipe as the 14B sibling, one scale down: a HybridSFT → HybridRL agent that alternates latent thoughts and text inside one generation stream, with an inspectable latent trace. The 8B checkpoint is the interesting one for local work — its GAIA average (43.7) is reported above DeepSeek-V4.1-Flash and Kimi-K3 in the paper’s own harness while running in ~16 GB bf16, and it beats every matched 8B text/latent baseline (GRPO, ARPO, Coconut, SoftThinking, Qwen3-8B thinking). Limits: no SWE-bench/Aider/LiveCodeBench/BFCL lane at all, no independent reproduction (Artificial Analysis has no entry), the frontier-baseline comparisons come from the authors’ own harness, and drop-in usability is low without the custom runtime. ⚠ Treat the GAIA/WebWalker deltas as lab-reported until a third party reruns them.

See also

Graph View

Backlinks