SmolLM3-3B at launch
Independent boards have not scored it on enough tests for the LLMBenchmark Index yet; below are the results HuggingFaceTB reported and the independent scores published so far. HuggingFaceTB's own announcement reports 16 benchmark results, including GPQA Diamond (41.7%) and Berkeley Function Calling Leaderboard (BFCL) (88.8).
What HuggingFaceTB reported at launch
From HuggingFaceTB's announcement. The maker's own numbers, not re-run by us.
-
AIME 2025
36.7%
Instruction Model; Extended Thinking
#11 of 11 models with a reported result
-
AIME 2025
9.3%
Instruction Model; No Extended Thinking
-
Berkeley Function Calling Leaderboard (BFCL)
88.8
Instruction Model; Extended Thinking
-
Berkeley Function Calling Leaderboard (BFCL)
92.3
Instruction Model; No Extended Thinking
-
Global MMLU
64.1%
Instruction Model; Extended Thinking
-
Global MMLU
53.5%
Instruction Model; No Extended Thinking
-
GPQA Diamond
41.7%
Instruction Model; Extended Thinking
#50 of 50 models with a reported result
-
GPQA Diamond
35.7%
Instruction Model; No Extended Thinking
-
GSM-Plus
83.4
Instruction Model; Extended Thinking
-
GSM-Plus
72.8
Instruction Model; No Extended Thinking
-
IFEval
71.2
Instruction Model; Extended Thinking
#13 of 13 models with a reported result
-
IFEval
76.7
Instruction Model; No Extended Thinking
Show all 16
-
LiveCodeBench v4
30
Instruction Model; Extended Thinking
-
LiveCodeBench v4
15.2
Instruction Model; No Extended Thinking
-
MixEval Hard
30.8
Instruction Model; Extended Thinking
-
MixEval Hard
26.9
Instruction Model; No Extended Thinking
Independent results so far
Not published