Benchmarks / BM-003
BM-003 · TESTED 2026-07-27
Qwen3-Coder-30B-A3B quantization sweep: Q4_K_M vs. Q6_K vs. Q8_0 on Arc B60 Pro
On 24GB Arc, Q6_K is the sweet spot: quality recovers to within 2 points of Q8_0 on HumanEval while staying 38% faster and fitting comfortably. Q8_0 barely fits and leaves no room for context. Q4_K_M is the budget pick when VRAM is tight.
vulkan · q6-k · llama.cpp b3500
Configuration
- Artifact
- Qwen3-Coder-30B-A3B-Instruct (multiple GGUF quants)
- Author
- Edgar
- Backend
- vulkan
- Checksum
- Context length
- 8192 tokens
- Model
- Qwen3-Coder-30B-A3B-Instruct
- Quantization
- q6-k
- Runtime
- llama.cpp b3500
- Settings
- -ngl 99 -c 8192 -t 8 -b 512
- Workload pack
- workload-pack-v1
Results
Identical workload pack run three times, one per quantization.
| MACHINE | PROMPT tok/s | GEN tok/s | TTFT ms | VRAM GB | RAM GB | POWER W |
|---|---|---|---|---|---|---|
| Jitori PC Q6_K quant. Fits with ~1.4GB headroom. | 398.1 | 27.4 | 980 | 22.6 | 3.4 | 250 |
| Jitori PC Q8_0 quant. Saturates VRAM; no context headroom. | 380.5 | 19.8 | 1320 | 24.0 | 4.1 | 252 |
| Jitori PC Q4_K_M quant. From bm-001, included for direct comparison. | 412.3 | 38.6 | 740 | 18.1 | 3.2 | 245 |
Quality scores
coding
0.81 / 0-1
HumanEval pass@1 (Q4_K_M)
coding
0.84 / 0-1
HumanEval pass@1 (Q6_K)
coding
0.86 / 0-1
HumanEval pass@1 (Q8_0)
Limitations
- Quality deltas are within typical run-to-run variance for single-pass HumanEval; treat as directional, not definitive.
- VRAM figures include KV cache at -c 8192; longer contexts shift the Q6_K vs Q8_0 boundary.
- Single machine (Jitori) — backend effects on other Arc SKUs may differ.
RELATED EQUIPMENT GUIDES
If these numbers point you toward hardware, these buyer-intent guides go deeper. Links below are editorial; they carry no affiliate tag.
-
The measured GPU buyer's guide
Turn this result into a hardware decision.
-
How much VRAM you actually need
Will your target model fit?
-
Build a rig around these numbers
A complete sub-$1,000 local-AI lab.