Aircade retro glove controller header
← All models

AEON QWEN

AEON QWEN model image

AEON QWEN

credit: AEON-7
input tok/s2,057.5
decode tok/s218.3
per-agent tok/s29.9
agent Pareto8 · NVIDIA GB10

NVFP4 MTP-XS · vLLM + z-lab DFlash speculative decode · selected fleet 8 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
4 builds
Robot-Filled Maze Shooter
4 builds
Ribbit Rush
0 builds
Builds
8 recorded outputs
Method
Pareto fleet 8
Runtime
vLLM + z-lab DFlash speculative decode
Quant
NVFP4 MTP-XS
Bench
8 agents / 218.3 decode tok/sec
benchmark sheet

Throughput profile

input tok/s2,057.5
decode tok/s218.3
per-agent tok/s29.9
agent Pareto8 · NVIDIA GB10
aggregate decode tok/secper-agent tok/sechighlight = selected fleet
1 agents
4.4 dec4.4/agent · 12.5 in
2 agents
100.7 dec52.0/agent · 915.5 in
4 agents
165.6 dec47.5/agent · 2,417.4 in
8 agents
218.3 dec29.9/agent · 2,057.5 in
16 agents
232.6 dec16.4/agent · 1,676.6 in
24 agents
251.1 dec11.4/agent · 1,581.6 in
32 agents
252.1 dec8.6/agent · 1,628.1 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
112.54.44.455.4155.4158.4
2915.5100.752.01.531.535.1
42417.4165.647.51.161.166.2
82057.5218.329.92.742.749.4
161676.6232.616.46.786.7817.6
241581.6251.111.410.7710.8124.5
321628.1252.18.612.5214.0232.5

Generated game outputs

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

Road Hopper 4 versions

road_hopper_aeon_teamwork · Road Hopper · teamwork evalfullscreen ↗ · compare
road-hopper_aeon_00 · Road Hopper · v1fullscreen ↗ · compare
road-hopper_aeon_01 · Road Hopper · v2 · brokenfullscreen ↗ · compare
road-hopper_aeon_02 · Road Hopper · v3 · brokenfullscreen ↗ · compare
Road Hopper teamwork eval AEON QWEN road_hopper_aeon_teamworkfullscreen ↗
Road Hopper v1 AEON QWEN road-hopper_aeon_00fullscreen ↗
Road Hopper v2 · broken AEON QWEN road-hopper_aeon_01fullscreen ↗
Road Hopper v3 · broken AEON QWEN road-hopper_aeon_02fullscreen ↗

Robot-Filled Maze Shooter 4 versions

maze_aeon_00 · Robot-Filled Maze Shooter · v1fullscreen ↗ · compare
maze_aeon_01 · Robot-Filled Maze Shooter · v2fullscreen ↗ · compare
maze_aeon_02 · Robot-Filled Maze Shooter · v3fullscreen ↗ · compare
maze_aeon_teamwork · Robot-Filled Maze Shooter · teamwork evalfullscreen ↗ · compare
Robot-Filled Maze Shooter v1 AEON QWEN maze_aeon_00fullscreen ↗
Robot-Filled Maze Shooter v2 AEON QWEN maze_aeon_01fullscreen ↗
Robot-Filled Maze Shooter v3 AEON QWEN maze_aeon_02fullscreen ↗
Robot-Filled Maze Shooter teamwork eval AEON QWEN maze_aeon_teamworkfullscreen ↗
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.