Inference
EN

Granite 4.2 3B vs Granite 4.2 8B

Change models
Granite 4.2 3BIBM
Granite 4.2 8BIBM
IndexAll categories, out of 100.
33.2 ±5.8
37.6 ±5.4 best
GapIndex points below the top.
−4.4 · no clear gap
Top
CoverageIndex weight that has results.
85%
85%
SpeedOutput tokens per second.
—
101/s
EloLMArena rating and rank.
1293
1319 best
ReleasedMonth of release.
2026-08
2026-08
Input / 1MUS dollars per 1M input tokens.
—
$0.06
Output / 1MUS dollars per 1M output tokens.
—
$0.15
1K chats2K tokens in, 500 out, each.
—
$0.195
1K documents20K tokens in, 1K out, each.
—
$1.35
1K agent tasks200K tokens in, 20K out, each.
—
$15.00
ContextMaximum tokens in one request.
128K
131K best
AgenticMulti-step tasks with tools.
33.3
36.8 best
CodingCode writing and repair.
34.9
36.5 best
ReasoningLogic problems and puzzles.
30.4
37.8 best
MultimodalTasks with images and text.
—
—
KnowledgeFacts and expert knowledge.
27.8
33.3 best
MultilingualTasks in many languages.
—
—
InstructionTasks with strict prompt rules.
58.5
63.2 best
MathMath problems.
34.3
41.9 best

Benchmarks

17 shared
Only Granite 4.2 8B: 3
Granite 4.2 3B
Granite 4.2 8B
Artificial Analysis Humanity's Last ExamKnowledge
6.6%166 / 222
9.7% best148 / 222
Artificial Analysis Long Context ReasoningReasoning
24.3%190 / 217
45.0% best164 / 217
Critical Physics TasksReasoning
0.0%150 / 214
0.3% best137 / 214
Artificial Analysis GPQA DiamondKnowledge
55.9%166 / 199
63.1% best150 / 199
Artificial Analysis Omniscience AccuracyKnowledge
9.2%163 / 164
11.8% best159 / 164
Graduate-Level Google-Proof Q&AKnowledge
54.8%119 / 141
64.1% best106 / 141
GDPval-AA normalizedAgentic
0.0%101 / 133
0.0%101 / 133
Artificial Analysis SciCodeCoding
25.3%108 / 114
31.5% best104 / 114
Massive Multitask Language Understanding ProfessionalKnowledge
67.8%60 / 75
74.0% best53 / 75
LiveCodeBench v6Coding
69.7%41 / 59
73.2% best38 / 59
Instruction Following BenchmarkInstruction
74.3%32 / 57
79.3% best16 / 57
American Invitational Mathematics Examination 2025Math
78.3%30 / 53
86.7% best24 / 53
Scientific Code BenchmarkCoding
24.1%35 / 35
36.1% best27 / 35
τ³-Bench Tool-Agent-User EvaluationAgentic
45.8%18 / 32
58.1% best16 / 32
Berkeley Function Calling Leaderboard v4Agentic
52.4%16 / 26
52.4%16 / 26
Harvard-MIT Mathematics Tournament February 2025Math
66.7%15 / 15
78.3% best13 / 15
Artificial Analysis Openness IndexKnowledge
72.2%
72.2%