LFM3 1.7B

Name: LFM3 1.7B
Author: Liquid AI

Liquid AI · LFM3 · released 2026-10-27 · data checked 2026-11-01

LFM3 1.7B is a 1.7B-parameter dense model from Liquid AI (United States), released 2026-10-27 under the LFM Open License license. It scores 60.0 on the Small LM Index, #4 of 21 tiny models. It needs about 1.7 GB of memory at 4-bit with an 8K-token context.

Hybrid convolution/attention model post-trained for agent harnesses; 256K context.

Small LM Index

60.0

#4 of 21 scored tiny models

Likely 58.2–61.8 · from 4 of 6 core benchmarks

Parameters

1.7B

Dense

Price per 1M tokens

—

No hosted API on OpenRouter; self-host

Context window

262K

tokens

Benchmarks

Each bar shows this model's score. The tick marks the best tiny model on that benchmark. Sources: 10 reported by Liquid AI. How we label sources.

GPQA Diamond Lab-reported Knowledge63.4

Class best: Nanbeige4.2 3B 87.4

Humanity's Last Exam Lab-reported Reasoning7.2

Class best: Spark-X2.6 4B 17.9

LiveCodeBench Lab-reported Coding · LiveCodeBench v673.5

SWE-bench Verified Lab-reported Coding · OpenCode harness43.4

Class best: Nanbeige4.2 3B 63.6

Terminal-Bench Lab-reported Agentic · Terminal-Bench 2.1 (OpenCode)12.9

Class best: Nanbeige4.2 3B 44.1

IFBench Lab-reported Instruction following73.0

Class best: Spark-X2.6 4B 75.8

IFEval Lab-reported Instruction following96.7

BFCL Lab-reported Agentic · BFCL v471.4

Class best: Spark-X2.6 4B 71.8

AA-LCR Lab-reported Long context52.8

Class best: Spark-X2.6 4B 67.3

SciCode Lab-reported Coding28.5

Class best: Spark-X2.6 4B 36.8

Not reported: estimated

What the index model expects from this model's other scores. Not a measurement.

MMLU-Pro≈ 79

AIME≈ 84

Specifications

Lab: Liquid AI (US · United States)
Architecture: Dense transformer
Total parameters: 1.7B
Active parameters: All (dense)
Size class: Tiny
Context window: 262K tokens
Input modalities: text
Reasoning: Hybrid (thinking on/off)
License: LFM Open License
Openness: Open weights
Memory to run: ~1.7 GB at 4-bit, 8K context (more sizes)
Hugging Face: 47.8K downloads (30 days) · 1.3K likes

Weights on Hugging Face ↗

Compare with

LFM3 1.7B vs Spark-X2.6 1.7B (1.7B, index 50.1)
LFM3 1.7B vs Qwen3.5 2B (2.3B, index 33.0)
LFM3 1.7B vs EXAONE 4.0 1.2B (1.3B, index 36.1)
LFM3 1.7B vs MiniCPM5 2B (2.5B, index 57.5)

Ranked in

Sources

https://huggingface.co/LiquidAI/LFM3-1.7B

Something wrong here? Report it.

Run it locally

Uses LiquidAI/LFM3-1.7B-GGUF (Q4_K_M). Swap the tag for another size: BF16, F16, Q4_0, Q4_K_M, Q5_K_M, Q6_K, Q8_0.

Ollama

ollama run hf.co/LiquidAI/LFM3-1.7B-GGUF:Q4_K_M

llama.cpp

llama-server -hf LiquidAI/LFM3-1.7B-GGUF:Q4_K_M

Then open http://localhost:8080 for a chat window.

MLX (Mac)

pip install -U mlx-lm
mlx_lm.chat --model LiquidAI/LFM3-1.7B-MLX-4bit

vLLM (GPU server)

vllm serve LiquidAI/LFM3-1.7B

Full-precision weights, for a server GPU.

Open in LM Studio

Memory needed: weights, context cache and runtime buffers.

Quantization	4K context	32K context	256K context
16-bit F16	4.0 GB	4.4 GB	7.2 GB
8-bit Q8_0	2.4 GB	2.8 GB	5.6 GB
4-bit Q4_K_M	1.6 GB	2.0 GB	4.8 GB

Check it against your GPU, Mac or phone →