Gemma-4-26B-A4B
Gemma-4-26B-A4B
credit: NVIDIA / Gemma
input tok/s6,803.0
decode tok/s778.2
per-agent tok/s33.9
agent Pareto32 · NVIDIA GB10
NVFP4 · vLLM + DFlash speculative decode · 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 + DFlash speculative decode
- Quant
- NVFP4
- Bench
- 32 agents / 778.2 decode tok/sec
benchmark sheet
Throughput profile
input tok/s6,803.0
decode tok/s778.2
per-agent tok/s33.9
agent Pareto32 · NVIDIA GB10
aggregate decode tok/secper-agent tok/sechighlight = selected fleet
1 agents227.5 dec227.6/agent · 3,548.3 in
2 agents131.6 dec96.9/agent · 1,462.9 in
4 agents401.6 dec128.0/agent · 6,262.5 in
8 agents482.3 dec83.8/agent · 6,432.1 in
16 agents495.8 dec48.4/agent · 6,383.4 in
24 agents743.1 dec40.6/agent · 6,507.1 in
32 agents778.2 dec33.9/agent · 6,803.0 in
48 agents741.2 dec20.9/agent · 5,903.3 in
64 agents807.3 dec16.3/agent · 6,449.4 in
Highlighted row is the selected build fleet / Pareto knee.
Raw sweep table
| agents | prompt_t/s | agg_gen_t/s | per_agent_gen_t/s | ttft_p50_s | ttft_p99_s | wall_s |
|---|---|---|---|---|---|---|
| 1 | 3548.3 | 227.5 | 227.6 | 0.20 | 0.20 | 1.1 |
| 2 | 1462.9 | 131.6 | 96.9 | 0.96 | 0.96 | 3.9 |
| 4 | 6262.5 | 401.6 | 128.0 | 0.45 | 0.45 | 2.5 |
| 8 | 6432.1 | 482.3 | 83.8 | 0.88 | 0.88 | 4.2 |
| 16 | 6383.4 | 495.8 | 48.4 | 1.78 | 1.78 | 8.3 |
| 24 | 6507.1 | 743.1 | 40.6 | 2.63 | 2.63 | 8.3 |
| 32 | 6803.0 | 778.2 | 33.9 | 3.35 | 3.36 | 10.5 |
| 48 | 5903.3 | 741.2 | 20.9 | 5.48 | 5.81 | 16.6 |
| 64 | 6449.4 | 807.3 | 16.3 | 5.52 | 7.09 | 20.3 |
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 ↗.
| Game | Score | Seed → Final | Assertions | Rounds | Seed → Final size | Build time |
|---|---|---|---|---|---|---|
| Road Hopper | 61.0 | 61.0 → 61.0 | 6/6 | 4 | 9 → 20 KB | 25 min |
| Robot-Filled Maze Shooter | 71.0 | 71.0 → 71.0 | 7/7 | 3 | 14 → 25 KB | 21 min |
Generated game outputs
Versions are listed first for selection. Embedded outputs remain below for direct review.
Road Hopper 1 versions
road_hopper_gemma_teamwork · Road Hopper · team buildfullscreen ↗ · compareRoad Hopper team build Gemma-4-26B
road_hopper_gemma_teamworkfullscreen ↗Robot-Filled Maze Shooter 1 versions
maze_gemma_teamwork · Robot-Filled Maze Shooter · team buildfullscreen ↗ · compareRobot-Filled Maze Shooter team build Gemma-4-26B
maze_gemma_teamworkfullscreen ↗