Aircade retro glove controller header
← All models

nvidia/Qwen3.6-35B-A3B-NVFP4

Qwen3.6-35B NVFP4 model image

nvidia/Qwen3.6-35B-A3B-NVFP4

credit: NVIDIA
input tok/s174.6
decode tok/s660.3
per-agent tok/s20.6
agent Pareto32 · NVIDIA GB10

NVFP4 (W4A16) + FP8 attn · vLLM 0.23.0, modelopt NVFP4+FP8, MARLIN MoE · context-calibrated team of 32 · 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 32 / 20-min build
Runtime
vLLM 0.23.0, modelopt NVFP4+FP8, MARLIN MoE
Quant
NVFP4 (W4A16) + FP8 attn
Bench
32 agents / 660.3 decode tok/sec
benchmark sheet

Throughput profile

input tok/s174.6
decode tok/s660.3
per-agent tok/s20.6
agent Pareto32 · NVIDIA GB10
aggregate decode tok/secper-agent tok/sechighlight = selected fleet
1 agents
77.3 dec77.3/agent · 20.2 in
2 agents
123.4 dec61.7/agent · 32.3 in
4 agents
194.4 dec48.7/agent · 50.9 in
8 agents
337.5 dec42.2/agent · 88.3 in
16 agents
489.3 dec30.6/agent · 128.8 in
32 agents
660.3 dec20.6/agent · 174.6 in

Highlighted row is the selected build fleet / Pareto knee.

Raw sweep table
agentsprompt_t/sagg_gen_t/sper_agent_gen_t/swall_s
120.277.377.33.31
232.3123.461.74.15
450.9194.448.75.27
888.3337.542.26.07
16128.8489.330.68.37
32174.6660.320.612.41
production sheet

Teamwork build

No fan-out: a 32-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 → 61.06/6413 → 21 KB21 min
Robot-Filled Maze Shooter71.061.0 → 71.06/7317 → 33 KB23 min

Generated game outputs

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

Road Hopper 1 versions

road_hopper_qwen36_35b_nvfp4_teamwork · Road Hopper · team buildfullscreen ↗ · compare
Road Hopper team build Qwen3.6-35B NVFP4 road_hopper_qwen36_35b_nvfp4_teamworkfullscreen ↗

Robot-Filled Maze Shooter 1 versions

maze_qwen36_35b_nvfp4_teamwork · Robot-Filled Maze Shooter · team buildfullscreen ↗ · compare
Robot-Filled Maze Shooter team build Qwen3.6-35B NVFP4 maze_qwen36_35b_nvfp4_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.