Aircade retro glove controller header
← All models

Laguna XS 2.1

Laguna XS 2.1 model image

Laguna XS 2.1

credit: Poolside / Laguna
B70 input tok/s66.3
B70 decode tok/s283.1
B70 per-agent tok/s10.6
B70 live fleet32 · Intel Arc Pro B70

Q4_K_M GGUF · llama.cpp/OpenCode on Atlas via local Ollama · single local model run with raw responses, validator output, real byte counts, and per-game telemetry recorded.

model dossier

Model sheet

Playable artifacts and benchmark context for this model page.

Road Hopper
1 builds
Robot-Filled Maze Shooter
2 builds
Ribbit Rush
0 builds
Builds
3 recorded outputs
Method
single local model run
Runtime
llama.cpp/OpenCode on Atlas via local Ollama
Quant
Q4_K_M GGUF
Bench
B70 live: 32 agents / 283.1 decode tok/sec
benchmark sheet

B70 live throughput profile

B70 input tok/s66.3
B70 decode tok/s283.1
B70 per-agent tok/s10.6
B70 live fleet32 · Intel Arc Pro B70
aggregate decode tok/secper-agent tok/sechighlight = selected live fleet
tool · 1 agents
63.7 dec88.4/agent · 339.0 in
tool · 8 agents
136.0 dec22.5/agent · 131.0 in
tool · 16 agents
183.1 dec13.3/agent · 106.2 in
tool · 32 agents
283.1 dec10.6/agent · 66.3 in
structured · 32 agents
161.4 dec6.4/agent · 91.3 in
novel · 32 agents
103.8 dec4.5/agent · 108.2 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
tool1339.063.788.41.41OK
tool8131.0136.022.55.95OK
tool16106.2183.113.38.91OK
tool3266.3283.110.611.87OK
structured3291.3161.46.442.43OK
novel32108.2103.84.566.54OK

Generated game outputs

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

Road Hopper 1 versions

road_hopper_laguna_xs21_q4_b70 · Road Hopper · Laguna XS 2.1 Q4_K_M B70 live teamwork evalfullscreen ↗ · compare
Road Hopper Laguna XS 2.1 Q4_K_M B70 live teamwork eval Laguna XS 2.1 road_hopper_laguna_xs21_q4_b70fullscreen ↗

Robot-Filled Maze Shooter 2 versions

maze_laguna_01 · Robot-Filled Maze Shooter · v2fullscreen ↗ · compare
maze_laguna_xs21_q4_b70 · Robot-Filled Maze Shooter · Laguna XS 2.1 Q4_K_M B70 live teamwork evalfullscreen ↗ · compare
Robot-Filled Maze Shooter v2 Laguna XS 2.1 maze_laguna_01fullscreen ↗
Robot-Filled Maze Shooter Laguna XS 2.1 Q4_K_M B70 live teamwork eval Laguna XS 2.1 maze_laguna_xs21_q4_b70fullscreen ↗
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.