Inference
EN

DeepSeek V3 vs LFM2.5-8B-A1B

Change models
DeepSeek V3DeepSeek
LFM2.5-8B-A1BLiquidAI
IndexAll categories, out of 100.
29.9 ±4.9
30.6 ±7.5 best
GapIndex points below the top.
−0.7 · no clear gap
Top
CoverageIndex weight that has results.
85% best
70%
SpeedOutput tokens per second.
37/s
335/s best
EloLMArena rating and rank.
1333
—
ReleasedMonth of release.
2024-12
2026-05
Input / 1MUS dollars per 1M input tokens.
$0.27
—
Output / 1MUS dollars per 1M output tokens.
$1.10
—
1K chats2K tokens in, 500 out, each.
$1.09
—
1K documents20K tokens in, 1K out, each.
$6.50
—
1K agent tasks200K tokens in, 20K out, each.
$76.00
—
ContextMaximum tokens in one request.
128K
128K
AgenticMulti-step tasks with tools.
26.0
46.3 best
CodingCode writing and repair.
31.1
—
ReasoningLogic problems and puzzles.
34.1 best
22.1
MultimodalTasks with images and text.
—
—
KnowledgeFacts and expert knowledge.
34.3 best
26.0
MultilingualTasks in many languages.
—
5.0
InstructionTasks with strict prompt rules.
29.1
44.8 best
MathMath problems.
27.2 best
27.1

Benchmarks

9 shared
Only DeepSeek V3: 25Only LFM2.5-8B-A1B: 12
DeepSeek V3
LFM2.5-8B-A1B
Artificial Analysis Humanity's Last ExamKnowledge
2.9%220 / 222
6.9% best163 / 222
Artificial Analysis Long Context ReasoningReasoning
29.3% best183 / 217
0.0%204 / 217
Critical Physics TasksReasoning
0.0%150 / 214
0.0%150 / 214
Artificial Analysis GPQA DiamondKnowledge
55.7% best168 / 199
51.3%171 / 199
Artificial Analysis Omniscience AccuracyKnowledge
25.5% best103 / 164
9.4%162 / 164
τ²-Bench Tool-Agent-User EvaluationAgentic
22.8%126 / 149
88.1% best47 / 149
Artificial Analysis IFBenchInstruction
34.8%122 / 146
55.6% best70 / 146
Instruction-Following EvalInstruction
86.1%36 / 70
91.8% best22 / 70
MATH-500 Problem SetMath
90.2% best21 / 27
88.8%23 / 27