Qwen3.8 27B MLX 4-bit on TensorFold, one Spark, thinking off
1× DGX Spark · TensorFold ·
Tested configuration
- Hardware
- 1x NVIDIA DGX Spark (128GB)
- Serving runtime
- TensorFold
- Runtime version or build
- TensorFold v0.6.0 with the recipe's five patches (image tag tensorfold-qwen38-27b:v0.6.0, patch hash fb91a0019fdc), built locally on nvcr.io/nvidia/pytorch:26.07-py3
- Weight quantization or precision
- MLX 4-bit
- Benchmark date
- Recipe credit
- MiaAI-Lab Qwen3.8 27B TensorFold recipeBy MiaAI-Lab
c11ff138871cf43d12898d79dbfa4dc952d17ca4
Decode sweep
~256 input / 512 output tokens per request. Output throughput is aggregate across concurrent requests. TTFT means time to first token.
| Concurrency | Output tok/s | TTFT p95 (s) |
|---|---|---|
| 1 | 50.25 | 0.44 |
| 2 | 89.48 | 0.78 |
| 4 | 128.82 | 1.38 |
| 8 | 179.13 | 2.86 |
| 12 | 197.68 | 19.64 |
| 16 | 201.49 | 25.03 |
Prefill and context
- Configured request window
- 262,144 tokens
- Longest tested prompt
- 245,760 target input tokens
- Validated retrieval
- ≈240K input tokens · passkey passed
Input throughput uses full request wall time. The passkey measures retrieval from one long prompt.
| Workload (c1) | Input tok/s | TTFT p95 (s) |
|---|---|---|
| 4,096 input tokens | 1,015.32 | 2.48 |
| 32,768 input tokens | 1,478.42 | 20.58 |
| 131,072 input tokens | 1,049.04 | 122.39 |
| 245,760 input tokensPasskey prompt | 787.01 | 311.47 |
Sustained load
20-minute soak at concurrency 8 · ~1024 input / 512 output tokens per request.
- Aggregate throughput
- 160.59 output tok/s
- TTFT p95
- 2.35 s
- Completed requests
- 382
- Request errors
- 0
Serving regression checks 13/13 passed
Tool-use evaluation
Incomplete pilot · 98.6% completionSafety gate failed
Comparisons require complete results with 100% completion under the same protocol ID. This run is retained as individual evidence.
- Protocol ID
tool-eval-standard-v1-b60e2f0- Workload
- 69 mock-tool scenarios per trial · 3 trials
- Mean score
- 90.0 / 100
- Pass@3 · passed at least once
- 85.3%
- Pass³ · passed every trial
- 85.3%