Benchmark Results
Seven-model comparison transcribed from the supplied image, including parameter counts, Average and 30 benchmark rows. Spark-X2.6-1.7B is highlighted; the final 4B column is a reference. Evaluation notes appear below the table.
Benchmark Results
Benchmark
Spark
X2.6-1.7B
Spark
X2.5-1.7B
Qwen3.5
2B
Gemma4
E2B
MiniCPM5
2B
LFM2.5-
2.6B
Spark
X2.6-4B
Total parameters
1.7B
1.7B
2.3B
5.1B (2.3B)
2.5B
2.7B
4.1B
Average
49.6
35.8
22.0
22.5
46.9
30.4
63.3
Agent
BFCL-V4
55.2
46.9
43.6*
36.6*
66.6*
61.1*
71.8
τ²-bench
85.3
65.3
48.8*
24.5*
82.2
69.7
86.9
τ³-bench Banking
21.3
20.1
4.1
8.8
20.8*
7.1
34.7
VitaBench2.0
22.6
8.3
5.2
4.4
3.4
3.7
30.8
MCP-Atlas
41.8
23.4
14.8
12.6
32.2
3.0
62.4
EnterpriseOps-Gym
26.1
10.5
6.8
8.6
14.8
-
33.6
ClawEval-EN
66.9
57.9
47.7
53.1*
58.4
62.9*
69.9
BrowseComp
41.7
29.7
3.1
3.7
23.4
16.6
53.1
SWebBench
53.9
31.6
22.8
15.6
45.7
-
71.7
PAgentBench
58.8
51.1
34.9
32.8
59.5
-
74.1
Code
SWE Pro
26.7
10.4
1.9
-
14.4*
0.6*
54.3
SWE Verified
59.8
28.3
6.8
-
46.4*
6.0*
62.8
SWE Multilingual
57.3
23.3
5.0
-
23.7
5.3
64.7
SciCode
20.1
18.2
6.0
20.5
26.3*
14.2*
36.8
Math
Gaokao 2026
118.9
114.8
94.0
81.8
107.9
93.8
134.9
AIME 2026
70.1
69.4
30.8
37.5*
86.5*
45.2*
91.2
HMMT Feb 26
48.8
48.4
21.5
20.5
63.8*
33.7*
81.3
IMO-AnswerBench
47.1
45.4
-
22.6
63.9
41.2
75.2
Long Context
AA-LCRv1.1
48.0
42.3
21.0*
16.3*
59.3*
5.7*
67.3
LongBench v2
40.4
33.2
38.7*
20.5
43.7*
30.3*
53.3
Graphwalks bfs 256k f1
85.1
28.6
60.2
42.5
71.1
27.0
92.2
Graphwalks bfs 1m f1
25.7
3.0
7.7
-
-
-
34.2
Openai_MRCR v2 256k-512k
56.1
4.8
16.9
-
-
-
65.6
Openai_MRCR v2 512k-1M
19.9
0.9
6.0
-
-
-
26.0
General & Knowledge
IFEval
89.4
89.5
78.6*
34.8
86.7*
93.4*
93.6
IFBench
65.5
66.3
41.3*
22.7
66.3*
59.2*
75.8
HLE(w/o tools)
8.7
6.3
2.1
2.5
8.9*
6.2*
17.9
HLE(w tools)
19.9
6.9
5.7
6.6
15.1
5.9
36.9
GPQA
51.3
43.8
44.6
43.8
55.1
60.7
69.9
GPQA Diamond
56.1
46.8
58.5
43.4*
70.2*
55.8*
77.1
•
* denotes reported results from publicly-released model cards / papers and - denotes scores not yet available. Spark-X2.6-4B is included as a reference.
•
All evaluations are conducted in thinking mode. Spark-X2.6-1.7B and Spark-X2.6-4B are evaluated with the thinking level set to high. The recommended sampling parameters for both
models 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.