Aircade retro glove controller header
← All models

Gemma-4-26B-A4B NVFP4 current plain vLLM

Gemma-4 current model image

Gemma-4-26B-A4B NVFP4 current plain vLLM

credit: NVIDIA / Gemma
input tok/s298.0
decode tok/s785.9
per-agent tok/s16.4
agent Pareto48 · NVIDIA GB10

NVFP4 · vLLM plain serve, 32k context, max 48 sequences · context-calibrated team of 5 · 20-min timed production loop  ·  Download on Hugging Face ↗.

model dossier

Model sheet

Playable artifacts and benchmark context for this model page.

Road Hopper
1 builds
Robot-Filled Maze Shooter
1 builds
Ribbit Rush
0 builds
Builds
2 recorded outputs
Method
team 5 / 20-min build
Runtime
vLLM plain serve, 32k context, max 48 sequences
Quant
NVFP4
Bench
48 agents / 785.9 decode tok/sec
benchmark sheet

Throughput profile

input tok/s298.0
decode tok/s785.9
per-agent tok/s16.4
agent Pareto48 · NVIDIA GB10
aggregate decode tok/secper-agent tok/sechighlight = selected fleet
1 agents
28.6 dec28.6/agent · 10.7 in
2 agents
58.4 dec29.3/agent · 21.9 in
4 agents
107.5 dec27.0/agent · 40.3 in
8 agents
185.4 dec23.2/agent · 69.5 in
16 agents
318.6 dec20.0/agent · 120.1 in
24 agents
457.1 dec19.1/agent · 172.8 in
32 agents
562.9 dec17.6/agent · 213.1 in
48 agents
785.9 dec16.4/agent · 298.0 in

Highlighted row is the selected build fleet / Pareto knee.

Raw sweep table
agentsprompt_t/sagg_gen_t/sper_agent_gen_t/swall_s
110.728.628.66.71
221.958.429.36.58
440.3107.527.07.14
869.5185.423.28.28
16120.1318.620.09.64
24172.8457.119.110.08
32213.1562.917.610.91
48298.0785.916.411.73
production sheet

Teamwork build

No fan-out: a 5-agent team collaborates on ONE game for a 20-minute timed loop — a lead assigns focus areas, workers improve in parallel, a merger integrates the best, validated each round so it never regresses. Team size comes from the context/teamwork Pareto (single-agent footprint × 1.10 → agents that fit the window). Methodology ↗.

GameScoreSeed → FinalAssertionsRoundsSeed → Final sizeBuild time
Road Hopper61.061.0 → 61.06/648 → 14 KB26 min
Robot-Filled Maze Shooter61.061.0 → 61.06/739 → 16 KB27 min

Generated game outputs

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

Road Hopper 1 versions

road_hopper_gemma4_nvfp4_current_teamwork · Road Hopper · team buildfullscreen ↗ · compare
Road Hopper team build Gemma-4 current road_hopper_gemma4_nvfp4_current_teamworkfullscreen ↗

Robot-Filled Maze Shooter 1 versions

maze_gemma4_nvfp4_current_teamwork · Robot-Filled Maze Shooter · team buildfullscreen ↗ · compare
Robot-Filled Maze Shooter team build Gemma-4 current maze_gemma4_nvfp4_current_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.