LFM2 2.6B

+ contribute

LFM2 2.6B is an open-weights language model from Liquid AI, released September 2025. Its published size is 2.57B parameters and a 33k-token context window. Licence: lfm1.0.

Catalogue identity notes

LiquidAI/LFM2-2.6B @36ed799 (created 2025-09-22, licence lfm1.0). Liquid AI LFM2 hybrid (conv + attention) instruct model, July-Sept 2025; earlier generation than the catalogued LFM2.5 models. Independent EuroEval rows (its own card prints only the 2.6B column next to uncredited comparators).

Model information sources

Release date ↗ · Weights licence ↗

HF safetensors metadata 2,569,272,320 parameters (LiquidAI/LFM2-2.6B @36ed799).

42.8±1.5#486 of 719 LLMsfrom 13 benchmark results across 2 sources

$ how it compares

The LLMs ranked either side of LFM2 2.6B, out of 828 with a score, then more from Liquid AI.

#	model	score	where	RAM	released
574	MiniCPM-V 2.6 OpenBMB	43±2	local	~6.4GB	2024-08
575	Qwen2.5 Coder 1.5B Instruct Alibaba	43±1	local	~2.4GB	2024-09
576	LFM2 2.6B Liquid AI	43±2	local	~3GB	2025-09
577	SmolLM3 3B Hugging Face	43±1	local	~3.3GB	2025-07
578	Granite 3.1 8B Instruct IBM	43±1	local	~6.6GB	2024-12
	more from Liquid AI
370	LFM2.5 2.6B Liquid AI	53±1	local	~3.2GB	2026-07
375	LFM2.5 8B-A1B Liquid AI	52±1	local	~6.5GB	2026-05
398	LFM2 24B-A2B Liquid AI	51±1	local	~16GB	2026-02

see all LLMs · find the ones that fit your machine

$ what it's good at

chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #243 of 393 * · 11% of its score · 1 test

Beats Mistral Small 4 (119B-A6.5B), Ministral 3 14B Reasoning 2512, Qwen2.5 Coder 32B Instruct here, though they rank higher overall. Trails Flan-T5 XXL, Granite 4.0 H 1B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
IFEval mean of four measures lab	79.6% ✎	█··▄▄·▄····▄ beats 4/5	easy · avg 85%	▇▇▇▇▇ 11%	typical

math▁▂▃▅▆▅▄█▄▁▁▁ #285 of 398 * · 11% of its score · 2 tests

Beats Llama 3.1 Nemotron 70B Instruct, GPT-4 (0613), Athene-70B here, though they rank higher overall. Trails DeepSeek R1 Distill Qwen 1.5B, Phi-2 (2.7B), StarCoder2 15B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
MATH-500 lab	63.6% ✎	▃····▃·█···▃ beats 1/5	easy · avg 81%	▇▇▇▇ 10%	at its level
GSM8K loose substring accuracy lab	82.4% ✎	█·█··███···█ beats 3/5	easy · avg 84%	▇ 1.0%	blunt

other tests10 not in a task · 77% of its score

test	result	vs the others in its run	difficulty	counts	why
MMLU lab	64.4% ✎	▄█···█·····▄ beats 3/5	easy · avg 74%	▇▇▇▇▇▇ 16%	at its level
MMLU-Pro lab	26.0% ✎	███···██···█ beats 2/5	medium · avg 45%	▇▇▇▇ 10%	typical
GPQA Diamond lab	26.6% ✎	██····███··█ beats 1/5	very hard · avg 35%	▇▇▇▇ 10%	sharp · too hard for it
IFBench (aggregation unspecified) lab	22.2% ✎	█··███··█··█ beats 2/5	hard · avg 33%	▇▇▇▇ 9.3%	typical
Multi-IF three-turn average lab	60.3% ✎	▄··█▄··▄···▄ beats 3/5	easy · avg 71%	▇▇▇▇ 9.3%	at its level
MMMLU seven languages (Liquid) lab	55.4% ✎	▄·█····█···▄ beats 4/5	medium · avg 60%	▇▇▇ 7.9%	at its level
MGSM six languages, loose (Liquid) lab	74.3% ✎	▄·▄·█···▄··▄ beats 3/5	easy · avg 82%	▇▇ 6.2%	at its level
EuroEval HellaSwag (English), test split, 5-shot generative MC, accuracy	64.5% ✎	▃·▁▁▂▁▂▅█▇▄▁ beats 39/100	medium · avg 68%	▇ 3.3%	at its level · noisy

2 more tests

test	result	vs the others in its run	difficulty	counts	why
EuroEval Life in the UK (English knowledge MC), test split, 5-shot, accuracy	76.3% ✎	▁▁▁▁▁▁▁▂▄█▄▃ beats 54/146	easy · avg 83%	▇ 2.8%	at its level · noisy
GSM-Plus lab	60.8% ✎	█·····███·██ beats 3/5	easy · avg 74%	▇ 1.9%	blunt · at its level

$ where to run it

llama.cpp · recommended

Install llama.cpp · Homebrew and winget installs name the commands llama-cli and llama-server instead.
macOS / Linux
curl -LsSf https://llama.app/install.sh | sh
Windows
irm https://llama.app/install.ps1 | iex
other ways to install ↗
Run LFM2 2.6B · needs ~1.56 GB · from huggingface.co/LiquidAI (official)
chat in the terminal
llama cli -hf LiquidAI/LFM2-2.6B-GGUF:Q4_K_M
Or start a local server and open the address it prints (http://localhost:9931)
llama serve -hf LiquidAI/LFM2-2.6B-GGUF:Q4_K_M

LM Studio

version

Install LM Studio · Open the app once before using the lms command.
download it (macOS / Windows / Linux) ↗
Run LFM2 2.6B · needs ~1.56 GB · from huggingface.co/LiquidAI (official)
open in LM Studio ↗
Or download it in the terminal
lms get https://huggingface.co/LiquidAI/LFM2-2.6B-GGUF@Q4_K_M -y
Then run lms ls to see its name, and lms chat followed by that name

Ollama

Install Ollama
download it (macOS / Windows) ↗
Linux
curl -fsSL https://ollama.com/install.sh | sh
Run LFM2 2.6B · needs ~1.56 GB · from huggingface.co/LiquidAI (official)
chat in the terminal
ollama run hf.co/LiquidAI/LFM2-2.6B-GGUF:Q4_K_M

MLX (Mac)

version

Install MLX (Mac) · Apple silicon Macs only, text only.
with uv
uv tool install mlx-lm
or
pip install mlx-lm
Run LFM2 2.6B · needs ~1.4 GB, Mac only · from huggingface.co/mlx-community
chat in the terminal
mlx_lm.chat --model mlx-community/LFM2-2.6B-4bit
Or start an OpenAI-compatible server at http://localhost:8080/v1
mlx_lm.server --model mlx-community/LFM2-2.6B-4bit

Official weights and documentation: huggingface.co/LiquidAI ↗

$ how the score was worked out · LFM2 Technical Report — Liqui… 93%, EuroEval leaderboards (result… 6%

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

source	results	share of the score	licence
LFM2 Technical Report — Liquid AI own SmolLM3 evaluation · lab-reported	11	▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 93%	CC BY 4.0
EuroEval leaderboards (results collection), English generative leaderboard	2	▇▇ 6%	MIT
prior · pull toward the middle		▇ 1%
weighted average		= 42.8

$ improve this page

Update model information Add a place to run it Add a test result Propose a benchmark Correct a benchmark score Edit data on GitHub

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated 2026-10-11 02:20 UTC · sources and licences · json