LiquidAI LFM3-1.7B at launch
- Liquid
- Released November 6, 2026
- 262K context
Liquid released LFM3-1.7B on November 6, 2026. Independent boards have not scored it on enough tests for the LLMBenchmark Index yet; below are the results Liquid reported and the independent scores published so far. Liquid's own announcement reports 27 benchmark results, including IFBench (73) and GPQA Diamond (63.4%). It costs $0.08 per million input tokens and $0.20 per million output tokens on OpenRouter.
What Liquid reported at launch
From “LFM3-1.7B”, November 6, 2026. The maker's own numbers, not re-run by us.
-
IFBench
73
#14 of 17 models with a reported result
-
AA-LCR
52.8
#9 of 11 models with a reported result
-
AA-Omniscience accuracy (public)
7.1
#2 of 2 models with a reported result
-
AA-Omniscience non-hallucination (public)
92.9
#1 of 2 models with a reported result
-
AA-Omniscience-Public
0.5
#2 of 2 models with a reported result
-
SciCode
28.5
#17 of 19 models with a reported result
-
AutomationBench (Public)
41.3
#2 of 6 models with a reported result
-
Benchmarks / Average
49.4
#1 of 2 models with a reported result
-
BFCL (v4)
71.4
#3 of 6 models with a reported result
-
BrowseComp-Plus (OpenClaw subset)
37.0%
#1 of 2 models with a reported result
-
EnterpriseOps-Gym
28.2%
#2 of 3 models with a reported result
-
Humanity's Last Exam
7.2%
no tools
Show all 27
-
Humanity's Last Exam
13.9%
with tools
#35 of 38 models with a reported result
-
GPQA Diamond
63.4%
#45 of 50 models with a reported result
-
IFEval
96.7
#2 of 13 models with a reported result
-
IFStruct
93.7
#1 of 2 models with a reported result
-
LiveClaw
45.6
#2 of 3 models with a reported result
-
LiveCodeBench Pro 25Q2
31.7
#1 of 2 models with a reported result
-
LiveCodeBench v6
73.5
#10 of 14 models with a reported result
-
MCP Atlas
32.6
#13 of 14 models with a reported result
-
Multi-IF
89.1
#1 of 5 models with a reported result
-
Ruler 128k
65.6
#2 of 3 models with a reported result
-
SWE-bench Multilingual (OpenCode)
23.0%
#1 of 2 models with a reported result
-
SWE-Bench Verified (OpenCode)
43.4%
#2 of 3 models with a reported result
-
Terminal-Bench 2.1 (OpenCode)
12.9%
#1 of 2 models with a reported result
-
ToolSandbox
80.4
#2 of 3 models with a reported result
-
WebArena Verified
38.4
#5 of 7 models with a reported result
Independent results so far
Not published