Benchmark Results
Six-model comparison transcribed from the latest supplied image, including Average and 30 benchmark rows. Spark-X2.6-4B is highlighted. Evaluation notes appear below the table.
Benchmark Results
Benchmark
Spark
X2.6-4B
Spark
X2.5-4B
Qwen3.5
9B
Qwen3.5
4B
Gemma4
12B
Gemma4
E4B
Average
63.3
52.1
49.1
43.2
47.0
29.7
Agent
BFCL-V4
71.8
65.1
66.1*
50.3*
37.4
47.0*
τ²-bench
86.9
75.1
79.1*
79.9*
69.0*
42.2*
τ³-bench
34.7
30.4
9.3
6.7
13.3
10.1
VitaBench2.0
30.8
25.2
15.6
18.2
12.4
4.8
MCP-Atlas
62.4
54.6
47.4*
40.8*
30.5*
15.0*
EnterpriseOps-Gym
33.6
27.3
24.7
22.9
24.5
8.6
ClawEval-EN
69.9
66.2
66.5*
62.3*
58.3
58.0*
BrowseComp
53.1
40.9
8.3
14.3
10.0
8.3
SWebBench
71.7
46.3
33.2
36.3
32.0
19.3
PAgentBench
74.1
59.4
53.8
51.4
51.8
37.0
Code
SWE Pro
54.3
44.4
33.8*
29.4*
21.9*
4.0*
SWE Verified
62.8
41.6
53.1*
38.8*
44.2*
14.0*
SWE Multilingual
64.7
53.3
43.3
27.7
32.5*
-
SciCode
36.8
34.7
32.7*
24.0
39.8
27.5
Math
Gaokao 2026
134.9
133.4
135.5
130.3
130.6
102.4
AIME 2026
91.2
90.7
88.2
83.0
82.1*
42.5*
HMMT Feb 26
81.3
81.2
70.8
69.7
65.6
34.2
IMO-AnswerBench
75.2
74.2
69.8
68.5
57.2
26.9
Long Context
AA-LCRv1.1
67.3
64.7
70.0*
63.0*
63.7*
32.0*
LongBench v2
53.3
48.7
55.2*
50.0*
45.3
27.8
Graphwalks bfs 256k f1
92.2
47.1
88.4
79.0
68.6
54.8
Graphwalks bfs 1m f1
34.2
2.5
25.3
15.8
-
-
Openai_MRCR v2 256k-512k
65.6
9.7
58.6
48.3
-
-
Openai_MRCR v2 512k-1M
26.0
2.9
34.3
12.2
-
-
General & Knowledge
IFEval
93.6
93.0
91.5*
89.8*
94.8
45.3
IFBench
75.8
75.0
64.5
59.2
73.5*
44.0*
HLE(w/o tools)
17.9
12.3
14.3
8.6
13.1
3.9
HLE(w tools)
36.9
21.8
17.9
20.2
19.0
10.1
GPQA
69.9
67.4
77.2
67.2
72.8
54.5
GPQA Diamond
77.1
73.7
81.7*
76.2*
78.8*
58.6*
•
* denotes reported results from publicly-released model cards / papers and - denotes scores not yet available.
•
All evaluations are conducted in thinking mode. Spark-X2.6-4B is evaluated with the thinking level set to high. The recommended sampling parameters for
Spark-X2.6-4B are temperature=1.0, top_p=0.95, and top_k=-1.
•
Gaokao 2026 consists of the five 2026 Chinese GAOKAO examinations (National I, National II, Beijing, Shanghai, Tianjin), each graded out of 150 points.
Dataset:
https://huggingface.co/datasets/XHToken/gaokao_math_2026
•
SWebBench: An in-house benchmark for web browsing, information retrieval, and research. It will be open-sourced soon.
•
PAgentBench (Personal Assistant Benchmark): An in-house benchmark for personal task management, file operations, and system interactions. It will be
open-sourced soon.