LFM3-1.7B

Liquid AI

LFM3-1.7B is Liquid AI's open-weight 1.73B text model for local agents and specialized applications, the text post-trained checkpoint of the LFM3 family released on October 27, 2026. It introduces a tri-hybrid architecture — 24 layers split between 13 gated short convolutions, 5 sliding-window attention layers with a 256-token window and 6 global attention layers — plus head-wise gated attention and partial RoPE, trained on 22T tokens. Weights are published under the LFM Open License v1.0 (commercial use allowed below a $10M annual revenue threshold) for llama.cpp, MLX, vLLM and ONNX runtimes.

Key Specifications

Parameters
1.7B
Context
-
Release Date
October 27, 2026
Average Score
50.7%

Timeline

Key dates in the model's history
Announcement
October 27, 2026
Last Update
November 3, 2026
Today
November 20, 2026

Technical Specifications

Parameters
1.7B
Training Tokens
22.0T tokens
Knowledge Cutoff
-
Family
-
Capabilities
MultimodalZeroEval

Benchmark Results

Model performance metrics across various tests and benchmarks

Programming

Programming skills tests
SWE-Bench Verified
Vendor table; OpenCode harness • Self-reported
43.4%

Reasoning

Logical reasoning and analysis
GPQA
Vendor table (GPQA Diamond) • Self-reported
63.4%

Other Tests

Specialized benchmarks
AA-LCR
Vendor table • Self-reported
52.8%
AutomationBench
Vendor table; AutomationBench (public set) • Self-reported
41.3%
BFCL-V4
Vendor table (BFCL V4) • Self-reported
71.4%
Claw-Eval
Vendor table; Claw-Eval General (EN) average • Self-reported
68.6%
Humanity's Last Exam
Vendor table; no tools • Self-reported
7.2%
Humanity's Last Exam (with tools, text-only)
Vendor table; with tools (search) • Self-reported
13.9%
IFBench
Vendor table • Self-reported
73.0%
IFEval
Vendor table • Self-reported
96.7%
LiveCodeBench v6
Vendor table • Self-reported
73.5%
MCP Atlas
Vendor table • Self-reported
32.6%
Multi-IF
Vendor table • Self-reported
89.1%
RULER 128k
Vendor table (RULER 128K) • Self-reported
65.6%
SciCode
Vendor table • Self-reported
28.5%
SWE-bench Multilingual
Vendor table; OpenCode harness • Self-reported
23.0%
TAU3-Bench
Vendor table (τ³-bench) • Self-reported
68.0%
Terminal-Bench 2.1
Vendor table; OpenCode harness • Self-reported
12.9%
WebArena-Verified
Vendor table • Self-reported
38.4%

License & Metadata

License
lfm1_0
Announcement Date
October 27, 2026
Last Updated
November 3, 2026

Similar Models

All Models

Recommendations are based on similarity of characteristics: developer organization, multimodality, parameter size, and benchmark performance. Choose a model to compare or go to the full catalog to browse all available AI models.