ToolSpace-35B-A3B

Code · Evaluation · 8B · 35B-A3B · 744B-A40B

ToolSpace-35B-A3B is a Qwen3.6-35B-A3B mixture-of-experts agent model trained with ToolSpace trajectories and on-policy agent reinforcement learning for long-horizon tool use.

ToolSpace constructs a tool-composition graph from real execution traces, capturing data, state, and retry dependencies. It synthesizes cross-environment tasks through state-aware walks with rollback and keeps trajectories that pass execution-based verification.

Paper: ToolSpace: Scaling Tool-Centric Environment Synthesis for Long-Horizon Agents.

Model information

Item	Details
Backbone	Qwen3.6-35B-A3B
Architecture	MoE, 35B total / 3B active parameters
Context length	262,144 tokens
Weight format	BF16 Safetensors
Training	Supervised fine-tuning + agent RL

Results

Results from Table 1 of the ToolSpace paper. Avg. is the unweighted mean over the seven benchmarks.

Model	AutomationBench	SkillsBench	JobBench	τ³-Bench Banking	BFCL V4	Harvey LAB	OfficeQA Pro	Avg.
Qwen3.6-35B-A3B	8.2	30.7	6.4	9.3	60.0	0.8	45.9	23.0
ToolSpace-35B-A3B (SFT)	45.8	42.4	40.0	34.0	74.3	1.7	48.9	41.0
ToolSpace-35B-A3B + Agent RL (this checkpoint)	45.8	45.3	40.0	42.3	74.8	2.5	51.9	43.2

Usage

Example vLLM deployment with four GPUs and a 64K serving window:

vllm serve dongguanting/ToolSpace-35B-A3B \
  --served-model-name ToolSpace-35B-A3B \
  --tensor-parallel-size 4 \
  --max-model-len 65536 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_xml \
  --reasoning-parser qwen3

Use the included tokenizer and chat template for text and tool-call messages.

Evaluation commands and benchmark configurations are available in the ToolSpace evaluation directory.