Qwen3-Coder-30B-A3B-Instruct BF16
Qwen3-Coder-30B-A3B-Instruct BF16
credit: Qwen
input tok/s87.0
decode tok/s344.2
per-agent tok/s10.8
agent Pareto32 · NVIDIA GB10
BF16 safetensors · vLLM BF16, 131k context for game builds · context-calibrated team of 6 · 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 6 / 20-min build
- Runtime
- vLLM BF16, 131k context for game builds
- Quant
- BF16 safetensors
- Bench
- 32 agents / 344.2 decode tok/sec
benchmark sheet
Throughput profile
input tok/s87.0
decode tok/s344.2
per-agent tok/s10.8
agent Pareto32 · NVIDIA GB10
aggregate decode tok/secper-agent tok/sechighlight = selected fleet
1 agents30.3 dec30.3/agent · 7.6 in
2 agents52.6 dec26.3/agent · 13.2 in
4 agents80.3 dec20.1/agent · 20.1 in
8 agents122.8 dec15.4/agent · 30.7 in
16 agents211.2 dec13.2/agent · 53.1 in
32 agents344.2 dec10.8/agent · 87.0 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 | wall_s |
|---|---|---|---|---|
| 1 | 7.6 | 30.3 | 30.3 | 8.46 |
| 2 | 13.2 | 52.6 | 26.3 | 9.73 |
| 4 | 20.1 | 80.3 | 20.1 | 12.75 |
| 8 | 30.7 | 122.8 | 15.4 | 16.67 |
| 16 | 53.1 | 211.2 | 13.2 | 19.39 |
| 32 | 87.0 | 344.2 | 10.8 | 23.80 |
Generated game outputs
Versions are listed first for selection. Embedded outputs remain below for direct review.
Road Hopper 1 versions
road_hopper_qwen3_coder_30b_bf16_teamwork · Road Hopper · team buildfullscreen ↗ · compareRoad Hopper team build Qwen3 Coder 30B
road_hopper_qwen3_coder_30b_bf16_teamworkfullscreen ↗Robot-Filled Maze Shooter 1 versions
maze_qwen3_coder_30b_bf16_teamwork · Robot-Filled Maze Shooter · team buildfullscreen ↗ · compareRobot-Filled Maze Shooter team build Qwen3 Coder 30B
maze_qwen3_coder_30b_bf16_teamworkfullscreen ↗