This is a single-machine, single-GB10 DGX Spark deployment of GLM-5.3-Flash using the NestQuant-1.5-4bit quantization. It does not use tensor parallelism or a second Spark.
| Item | Configuration |
|---|---|
| Hardware | 1 × NVIDIA DGX Spark, GB10 Grace Blackwell, 128 GB unified memory |
| Quantization | NestQuant 1.5/4-bit |
| Expert allocation | 2,630 active slots, 20% fixed |
| Predictor | 75% jT / 25% EMA, updated every token |
| Speculative decoding | MTP2 |
| Sampling | Temperature 0.7, top-p 0.95, thinking off |
| Context | 110,592 tokens; 1 GiB KV cache |
Verified LocalMaxxing measurement
| Decode | Prefill estimate | TTFT | MTP acceptance |
|---|---|---|---|
| 18.4 tok/s | 40.3 tok/s | 6.93 s | 83.33% (160/192) |
The prefill number is LocalMaxxing's end-to-end estimate from prompt tokens divided by TTFT. It includes scheduling, expert delivery, mailbox acknowledgement and the first decode step, so it is not an isolated layer-major prefill kernel measurement.
The attached run uses LocalMaxxing's canonical reasoning-v1 prompt and passed verified-run validation with no issues.
Reproduce
bash <(curl -fsSL https://raw.githubusercontent.com/jarrelscy/spark-bench/228ff83a27608189efb82349e4adf13bfb253e8f/scripts/reproduce_nestquant_single_spark.sh) --spark USER@SPARK_HOST --reasoning off --keep-serving
- Quantized model and single-Spark documentation
- Serving implementation and recipe
- Spark Bench result and transcripts
- Docker image:
docker.io/jarrelscy/nestquant@sha256:3859a4e0018ec915ca0d4a3f9488158026fc2525534051433e2b37a729102364
Attached evidenceLocalMaxxing runs referenced by this report — open any run to see the full submission