LFM2.5 350M

Name: LFM2.5 350M
Author: Liquid AI

Liquid AI · LFM2.5 · released 2026-03-31 · data checked 2026-11-01

LFM2.5 350M is a 350M-parameter dense model from Liquid AI (United States), released 2026-03-31 under the LFM Open License license. It scores 12.6 on the Small LM Index, #19 of 21 tiny models. It needs about 0.8 GB of memory at 4-bit with an 8K-token context.

Small LM Index

12.6

#19 of 21 scored tiny models

Likely 9.5–16.6 · from 2 of 6 core benchmarks

Parameters

350M

Dense

Price per 1M tokens

—

No hosted API on OpenRouter; self-host

Context window

32K

tokens

Benchmarks

Each bar shows this model's score. The tick marks the best tiny model on that benchmark. Sources: 6 reported by Liquid AI. How we label sources.

MMLU-Pro Lab-reported Knowledge20.0

Class best: Qwen3.5 4B 79.1

GPQA Diamond Lab-reported Knowledge30.6

Class best: Nanbeige4.2 3B 87.4

τ²-Bench Lab-reported Agentic · Average of Telecom and Retail18.4

Class best: MiniCPM5 2B 97.1

IFBench Lab-reported Instruction following40.7

Class best: Spark-X2.6 4B 75.8

IFEval Lab-reported Instruction following77.0

Class best: LFM3 1.7B 96.7

BFCL Lab-reported Agentic · BFCL v421.9

Class best: Spark-X2.6 4B 71.8

Not reported: estimated

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

Humanity's Last Exam≈ 1

AIME≈ 6

LiveCodeBench≈ 11

SWE-bench Verified≈ 3

Specifications

Lab: Liquid AI (US · United States)
Architecture: Dense transformer
Total parameters: 350M
Active parameters: All (dense)
Size class: Tiny
Context window: 32K tokens
Input modalities: text
Reasoning: Instruct (no thinking)
License: LFM Open License
Openness: Open weights
Memory to run: ~0.8 GB at 4-bit, 8K context (more sizes)
Hugging Face: 114K downloads (30 days) · 441 likes

Weights on Hugging Face ↗

Compare with

LFM2.5 350M vs Qwen3.5 0.8B (900M, index 10.7)
LFM2.5 350M vs Gemma 3 1B (1B, index 6.1)
LFM2.5 350M vs LFM2.5 1.2B Thinking (1.2B, index 30.1)
LFM2.5 350M vs LFM2.5 1.2B Instruct (1.2B, index 20.6)

Sources

https://huggingface.co/LiquidAI/LFM2.5-350M

Something wrong here? Report it.

Run it locally

Uses LiquidAI/LFM2.5-350M-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/LFM2.5-350M-GGUF:Q4_K_M

llama.cpp

llama-server -hf LiquidAI/LFM2.5-350M-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/LFM2.5-350M-MLX-4bit

vLLM (GPU server)

vllm serve LiquidAI/LFM2.5-350M

Full-precision weights, for a server GPU.

Open in LM Studio

Memory needed: weights, context cache and runtime buffers.

Quantization	4K context	32K context
16-bit F16	1.3 GB	1.6 GB
8-bit Q8_0	0.9 GB	1.3 GB
4-bit Q4_K_M	0.8 GB	1.1 GB

Check it against your GPU, Mac or phone →