Inference
EN
LiquidAI

LFM2.5-8B-A1B

LFM2.5-8B-A1B is a reasoning model from LiquidAI. 13 benchmarks count toward its index, in 6 categories.

  • Open weights
  • Reasoning
  • Input: text
  • Released 2026-05
RankedEnough results for an index.
IndexAll categories, out of 100.30.6±7.5Median 51.0
CoverageIndex weight that has results.70%
SpeedOutput tokens per second.335/sMedian 67/s
First tokenSeconds to the first output token.7.6 sMedian 1.3 s
Input / 1MUS dollars per 1M input tokens.—Median $0.60
Output / 1MUS dollars per 1M output tokens.—Median $2.55
ContextMaximum tokens in one request.128KMedian 262K
EloLMArena rating and rank.—Median 1422

Run locally

8.47B parameters
16-bit
18.8 GB
8-bit
10.4 GB
4-bit
6.4 GB
Weights
Hugging Face (opens in a new tab) ↗

Memory need with a short context. The 4-bit copy fits in 8 GB. Compare local models

Capabilities

Index per category, out of 100.
Out of 100
AgenticMulti-step tasks with tools.2 counted46.3159 / 218
CodingCode writing and repair.None counted——
ReasoningLogic problems and puzzles.2 counted22.1222 / 227
MultimodalTasks with images and text.None counted——
KnowledgeFacts and expert knowledge.3 counted26.0254 / 268
MultilingualTasks in many languages.1 counted5.0—
InstructionTasks with strict prompt rules.2 counted44.838 / 63
MathMath problems.3 counted27.1150 / 168

Results

13 counted
BenchmarkThe benchmark name.ResultThe score from the source.PlacePlace among results on this benchmark.
Instruction-Following EvalInstructionAbove fitFar above this model's index.91.8%22 / 70
τ²-Bench Tool-Agent-User EvaluationAgenticAbove fitFar above this model's index.88.1%47 / 149
Artificial Analysis IFBenchInstruction55.6%70 / 146
Berkeley Function Calling Leaderboard v4Agentic49.7%18 / 26
Critical Physics TasksReasoning0.0%150 / 214
Artificial Analysis Humanity's Last ExamKnowledge6.9%163 / 222
EuroEval SpanishMultilingual24.1% ±1.1130 / 176
EuroEval FrenchMultilingual29.6% ±0.8136 / 178
EuroEval GermanMultilingual20.6% ±0.9149 / 191
EuroEval ItalianMultilingual20.8% ±1.1136 / 174
Instruction Following BenchmarkInstruction56.5%47 / 57
EuroEval PortugueseMultilingual22.0% ±1.0152 / 182
American Invitational Mathematics Examination 2025Math42.5%45 / 53
MATH-500 Problem SetMath88.8%23 / 27
Artificial Analysis GPQA DiamondKnowledge51.3%171 / 199
EuroEval SwedishMultilingual14.3% ±1.2171 / 195
EuroEval DutchMultilingual21.1% ±0.7154 / 175
EuroEval PolishMultilingual10.8% ±0.8187 / 209
Artificial Analysis Long Context ReasoningReasoning0.0%204 / 217
AIME 2026Math50.0%32 / 34
Artificial Analysis Omniscience AccuracyKnowledge9.4%162 / 164

Sources

BenchLMCC BY-NC 4.0EuroEvalMIT

History

5 releases
Index31.5 → 30.6
23 Oct 202612 Nov 2026

Questions

How good is LFM2.5-8B-A1B?
Its index is 30.6 out of 100. That is place 248 of 283 ranked models.
What is LFM2.5-8B-A1B best at?
Instruction: 44.8 out of 100, place 38 of 63.
How fast is LFM2.5-8B-A1B?
335 output tokens per second, 7.6 s to the first token.
What is the context window of LFM2.5-8B-A1B?
128K tokens.
Is LFM2.5-8B-A1B open source?
Yes. Its weights are published.
When was LFM2.5-8B-A1B released?
May 2026.

Badge

LFM2.5-8B-A1B on Inference360: place 248 of 283, index 30.6
HTML
<a href="https://inference360.ai/models/lfm2-5-8b-a1b"><img src="https://inference360.ai/badges/lfm2-5-8b-a1b.svg" alt="LFM2.5-8B-A1B on Inference360: place 248 of 283, index 30.6" height="20"></a>
Markdown
[![LFM2.5-8B-A1B on Inference360: place 248 of 283, index 30.6](https://inference360.ai/badges/lfm2-5-8b-a1b.svg)](https://inference360.ai/models/lfm2-5-8b-a1b)

Compare with

Sarvam 30B30.3DeepSeek V329.9Magistral Small31.4LFM2.5-2.6B29.3

More from LiquidAI

All models
LFM2.5-1.2B-JP-202606UnrankedLFM2.5-VL-1.6B-ExtractUnrankedLFM2.5-VL-450M-ExtractUnrankedLFM2.5-ColBERT-350MUnrankedLFM2.5-230MUnranked