Aircade retro glove controller header
← All models

Nex-N2-mini

NEX2 model image

Nex-N2-mini

credit: nex-agi base / s-batman NVFP4 GGUF conversion
B70 input tok/s32.7
B70 decode tok/s224.3
B70 per-agent tok/s7.9
B70 live fleet32 · Intel Arc Pro B70

NVFP4 GGUF · llama.cpp · context-calibrated team of 3 · 20-min timed production loop  ·  Download on Hugging Face ↗.

model dossier

Model sheet

Playable artifacts and benchmark context for this model page.

Road Hopper
2 builds
Robot-Filled Maze Shooter
3 builds
Ribbit Rush
0 builds
Builds
5 recorded outputs
Method
team 3 / 20-min build
Runtime
llama.cpp
Quant
NVFP4 GGUF
Bench
B70 live: 32 agents / 224.3 decode tok/sec
benchmark sheet

B70 live throughput profile

B70 input tok/s32.7
B70 decode tok/s224.3
B70 per-agent tok/s7.9
B70 live fleet32 · Intel Arc Pro B70
aggregate decode tok/secper-agent tok/sechighlight = selected live fleet
tool · 1 agents
62.7 dec82.3/agent · 131.0 in
tool · 8 agents
142.3 dec19.8/agent · 92.8 in
tool · 16 agents
163.0 dec11.2/agent · 57.4 in
tool · 32 agents
224.3 dec7.9/agent · 32.7 in
structured · 32 agents
165.3 dec6.6/agent · 43.1 in
novel · 32 agents
81.9 dec5.5/agent · 55.3 in

Highlighted row is the selected 32-agent tool-prompt live fleet on Intel Arc Pro B70. Structured and novel rows are LocalMaxxing prompt classes.

Raw B70 live table
classagentsprefill_tpsagg_decode_tpsper_agent_tpsbatch_wall_sstatus
tool1131.062.782.34.09OK
tool892.8142.319.814.39OK
tool1657.4163.011.225.13OK
tool3232.7224.37.936.53OK
structured3243.1165.36.649.57OK
novel3255.381.95.529.35OK
production sheet

Teamwork build

No fan-out: a 3-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 → 61.06/6318 → 27 KB31 min
Robot-Filled Maze Shooter71.071.0 → 71.07/7128 → 32 KB20 min
maze_loop_hour— → —10 → 19 KB60 min

Generated game outputs

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

Road Hopper 2 versions

road_hopper_nex2_teamwork · Road Hopper · team buildfullscreen ↗ · compare
road_hopper_nex2_b70_live · Road Hopper · team buildfullscreen ↗ · compare
Road Hopper team build NEX2 road_hopper_nex2_teamworkfullscreen ↗
Road Hopper team build NEX2 road_hopper_nex2_b70_livefullscreen ↗

Robot-Filled Maze Shooter 3 versions

maze_nex2_b70_live · Robot-Filled Maze Shooter · team buildfullscreen ↗ · compare
maze_nex2_loop_hour · Robot-Filled Maze Shooter · team buildfullscreen ↗ · compare
maze_nex2_teamwork · Robot-Filled Maze Shooter · team buildfullscreen ↗ · compare
Robot-Filled Maze Shooter team build NEX2 maze_nex2_b70_livefullscreen ↗
Robot-Filled Maze Shooter team build NEX2 maze_nex2_loop_hourfullscreen ↗
Robot-Filled Maze Shooter team build NEX2 maze_nex2_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.