Qwen3.6-35B-A3B UD-Q5_K_XL GGUF
Qwen3.6-35B-A3B UD-Q5_K_XL GGUF
credit: Qwen / local Turbo
input tok/s61.8
decode tok/s158.5
per-agent tok/s16.3
agent Pareto12 · Intel Arc Pro B70
UD-Q5_K_XL GGUF · llama.cpp SYCL on Intel Arc Pro B70, 12-slot unified KV · selected fleet 12 agents (its Pareto knee). Full throughput curve on Benchmarks.
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
- Pareto fleet 12
- Runtime
- llama.cpp SYCL on Intel Arc Pro B70, 12-slot unified KV
- Quant
- UD-Q5_K_XL GGUF
- Bench
- 12 agents / 158.5 decode tok/sec
benchmark sheet
Throughput profile
input tok/s61.8
decode tok/s158.5
per-agent tok/s16.3
agent Pareto12 · Intel Arc Pro B70
aggregate decode tok/secper-agent tok/sechighlight = selected fleet
1 agents68.2 dec79.1/agent · 307.9 in
4 agents128.0 dec37.7/agent · 116.6 in
8 agents156.7 dec24.1/agent · 78.4 in
12 agents158.5 dec16.3/agent · 61.8 in
Highlighted row is the 12-slot B70 live serve ceiling.
Raw sweep table
| agents | prompt_t/s | agg_gen_t/s | per_agent_gen_t/s | wall_s |
|---|---|---|---|---|
| 1 | 307.93 | 68.15 | 79.12 | 3.76 |
| 4 | 116.59 | 127.97 | 37.70 | 8.00 |
| 8 | 78.41 | 156.72 | 24.10 | 13.07 |
| 12 | 61.75 | 158.48 | 16.31 | 19.38 |
Generated game outputs
Versions are listed first for selection. Embedded outputs remain below for direct review.
Road Hopper 1 versions
road-hopper_qwen36_35b_q5_turbo_00 · Road Hopper · v1fullscreen ↗ · compareRoad Hopper v1 Qwen3.6-35B Q5 Turbo
road-hopper_qwen36_35b_q5_turbo_00fullscreen ↗Robot-Filled Maze Shooter 1 versions
maze_qwen36_35b_q5_turbo_00 · Robot-Filled Maze Shooter · v1fullscreen ↗ · compareRobot-Filled Maze Shooter v1 Qwen3.6-35B Q5 Turbo
maze_qwen36_35b_q5_turbo_00fullscreen ↗