Spark Bench

Recursive Flint 40B-A2.5B NVFP4 GGUF: llama.cpp b11856, one slot (-np 1), one Spark, 262K

1× DGX Spark · llama.cpp ·

Tested configuration

Hardware
1x NVIDIA DGX Spark (128GB)
Serving runtime
llama.cpp
Runtime version or build
llama.cpp b11856 (3642238cc120, MIT), built locally from .devops/cuda.Dockerfile with CUDA 13.2 and CUDA architecture 121 (built as 121a). Nothing was pulled from a third-party registry and nothing was published
Weight quantization or precision
NVFP4 GGUF
Benchmark date
Recipe credit
RecursiveSI/flint-nvfp4-ggufBy RecursiveSIModel card llama-server command (RecursiveSI/flint-nvfp4-gguf); no recipe repository

Decode sweep

~256 input / 512 output tokens per request. Output throughput is aggregate across concurrent requests. TTFT means time to first token.

ConcurrencyOutput tok/sTTFT p95 (s)
181.730.32
283.406.46
482.8518.95
883.0343.81
1281.6769.43
1682.9593.22

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/sTTFT p95 (s)
4,096 input tokens1,485.242.00
32,768 input tokens1,848.5916.74
131,072 input tokens1,436.0489.89
245,760 input tokensPasskey prompt1,090.53220.19

Sustained load

20-minute soak at concurrency 8 · ~1024 input / 512 output tokens per request.

Aggregate throughput
74.62 output tok/s
TTFT p95
48.81 s
Completed requests
182
Request errors
0

Serving regression checks 10/13 passed

Tool-use evaluation

Complete · 100% completionSafety gate passed

Compare only against complete results with 100% completion under the same protocol ID.

Protocol ID
tool-eval-standard-v1-b60e2f0
Workload
69 mock-tool scenarios per trial · 3 trials
Mean score
86.0 / 100
Pass@3 · passed at least once
84.1%
Pass³ · passed every trial
75.4%

Read the tool-eval report and scenario findings