Aircade retro glove controller header
← All models

AgentWorld-35B-A3B

AgentWorld-35B model image

AgentWorld-35B-A3B

credit: Qwen / AgentWorld
input tok/s1,055.7
decode tok/s145.4
per-agent tok/s4.7
agent Pareto32 · NVIDIA GB10

GGUF-NVFP4 (qwen35moe) · llama.cpp · selected fleet 32 agents (its Pareto knee)  ·  Download on Hugging Face ↗. Full throughput curve on Benchmarks.

model dossier

Model sheet

Playable artifacts and benchmark context for this model page.

Road Hopper
12 builds
Robot-Filled Maze Shooter
11 builds
Ribbit Rush
0 builds
Builds
23 recorded outputs
Method
Pareto fleet 32
Runtime
llama.cpp
Quant
GGUF-NVFP4 (qwen35moe)
Bench
32 agents / 145.4 decode tok/sec
benchmark sheet

Throughput profile

input tok/s1,055.7
decode tok/s145.4
per-agent tok/s4.7
agent Pareto32 · NVIDIA GB10
aggregate decode tok/secper-agent tok/sechighlight = selected fleet
1 agents
46.8 dec46.8/agent · 1,015.6 in
2 agents
71.9 dec36.1/agent · 1,452.6 in
4 agents
99.4 dec25.0/agent · 1,398.8 in
8 agents
121.5 dec15.3/agent · 1,299.2 in
16 agents
132.2 dec8.4/agent · 1,230.8 in
24 agents
120.0 dec5.1/agent · 1,085.2 in
32 agents
145.4 dec4.7/agent · 1,055.7 in

Highlighted row is the selected build fleet / Pareto knee.

Raw sweep table
agentsprompt_t/sagg_gen_t/sper_agent_gen_t/sttft_p50_sttft_p99_swall_s
11015.646.846.80.680.685.5
21452.671.936.10.730.967.1
41398.899.425.01.952.0110.3
81299.2121.515.33.404.3416.9
161230.8132.28.45.989.2431.0
241085.2120.05.18.7615.7651.2
321055.7145.44.711.8721.6256.3

Generated game outputs

Versions are listed first for selection. Embedded outputs remain below for direct review.

Road Hopper 12 versions

road_hopper_agentworld_teamwork · Road Hopper · teamwork evalfullscreen ↗ · compare
road_hopper_b70_crewai_n12 · Road Hopper · CrewAI N=12fullscreen ↗ · compare
road_hopper_b70_solo · Road Hopper · Solofullscreen ↗ · compare
road-hopper_agentworld_00 · Road Hopper · v1fullscreen ↗ · compare
road-hopper_agentworld_01 · Road Hopper · v2fullscreen ↗ · compare
road-hopper_agentworld_02 · Road Hopper · v3fullscreen ↗ · compare
road_hopper_agentworld_b70_live · Road Hopper · B70 live teamwork evalfullscreen ↗ · compare
road_hopper_agentworld_b70_q5 · Road Hopper · B70 Q5_K_M quant evalfullscreen ↗ · compare
road_hopper_agentworld_b70_q6 · Road Hopper · B70 Q6_K quant eval run 1fullscreen ↗ · compare
road_hopper_agentworld_b70_q6b · Road Hopper · B70 Q6_K quant eval run 2fullscreen ↗ · compare
road_hopper_agentworld_b70_stock · Road Hopper · B70 stock baselinefullscreen ↗ · compare
road_hopper_b70_crewai_n6 · Road Hopper · CrewAI N=6 · brokenfullscreen ↗ · compare
Road Hopper teamwork eval AgentWorld-35B road_hopper_agentworld_teamworkfullscreen ↗
Road Hopper CrewAI N=12 AgentWorld-35B road_hopper_b70_crewai_n12fullscreen ↗
Road Hopper Solo AgentWorld-35B road_hopper_b70_solofullscreen ↗
Road Hopper v1 AgentWorld-35B road-hopper_agentworld_00fullscreen ↗
Road Hopper v2 AgentWorld-35B road-hopper_agentworld_01fullscreen ↗
Road Hopper v3 AgentWorld-35B road-hopper_agentworld_02fullscreen ↗
Road Hopper B70 live teamwork eval AgentWorld-35B road_hopper_agentworld_b70_livefullscreen ↗
Road Hopper B70 Q5_K_M quant eval AgentWorld-35B road_hopper_agentworld_b70_q5fullscreen ↗
Road Hopper B70 Q6_K quant eval run 1 AgentWorld-35B road_hopper_agentworld_b70_q6fullscreen ↗
Road Hopper B70 Q6_K quant eval run 2 AgentWorld-35B road_hopper_agentworld_b70_q6bfullscreen ↗
Road Hopper B70 stock baseline AgentWorld-35B road_hopper_agentworld_b70_stockfullscreen ↗
Road Hopper CrewAI N=6 · broken AgentWorld-35B road_hopper_b70_crewai_n6fullscreen ↗

Robot-Filled Maze Shooter 11 versions

maze_agentworld_00 · Robot-Filled Maze Shooter · v1fullscreen ↗ · compare
maze_agentworld_01 · Robot-Filled Maze Shooter · v2fullscreen ↗ · compare
maze_agentworld_02 · Robot-Filled Maze Shooter · v3fullscreen ↗ · compare
maze_agentworld_b70_live · Robot-Filled Maze Shooter · B70 live teamwork evalfullscreen ↗ · compare
maze_agentworld_b70_q5 · Robot-Filled Maze Shooter · B70 Q5_K_M quant evalfullscreen ↗ · compare
maze_agentworld_b70_q6 · Robot-Filled Maze Shooter · B70 Q6_K quant eval run 1fullscreen ↗ · compare
maze_agentworld_b70_q6b · Robot-Filled Maze Shooter · B70 Q6_K quant eval run 2fullscreen ↗ · compare
maze_agentworld_b70_stock · Robot-Filled Maze Shooter · B70 stock baselinefullscreen ↗ · compare
maze_agentworld_teamwork · Robot-Filled Maze Shooter · teamwork evalfullscreen ↗ · compare
maze_b70_crewai_n12 · Robot-Filled Maze Shooter · CrewAI N=12fullscreen ↗ · compare
maze_b70_solo · Robot-Filled Maze Shooter · Solofullscreen ↗ · compare
Robot-Filled Maze Shooter v1 AgentWorld-35B maze_agentworld_00fullscreen ↗
Robot-Filled Maze Shooter v2 AgentWorld-35B maze_agentworld_01fullscreen ↗
Robot-Filled Maze Shooter v3 AgentWorld-35B maze_agentworld_02fullscreen ↗
Robot-Filled Maze Shooter B70 live teamwork eval AgentWorld-35B maze_agentworld_b70_livefullscreen ↗
Robot-Filled Maze Shooter B70 Q5_K_M quant eval AgentWorld-35B maze_agentworld_b70_q5fullscreen ↗
Robot-Filled Maze Shooter B70 Q6_K quant eval run 1 AgentWorld-35B maze_agentworld_b70_q6fullscreen ↗
Robot-Filled Maze Shooter B70 Q6_K quant eval run 2 AgentWorld-35B maze_agentworld_b70_q6bfullscreen ↗
Robot-Filled Maze Shooter B70 stock baseline AgentWorld-35B maze_agentworld_b70_stockfullscreen ↗
Robot-Filled Maze Shooter teamwork eval AgentWorld-35B maze_agentworld_teamworkfullscreen ↗
Robot-Filled Maze Shooter CrewAI N=12 AgentWorld-35B maze_b70_crewai_n12fullscreen ↗
Robot-Filled Maze Shooter Solo AgentWorld-35B maze_b70_solofullscreen ↗
UNDER RAPID CONSTRUCTION - WILL CHANGE
why

Local agents need public test loops

Working on local agents matters because these tests are the first step toward an R&D base-model trust score: a practical signal for choosing which architectures and models deserve deeper research and development.

Agentic Arcade is a playable hello world for that work. Anyone can open the games, review the artifacts, and compare model behavior. The mission is maximal distribution of intelligence to all people.