LFM2.5 8B-A1B

+ contribute

LFM2.5 8B-A1B is an open-weights language model from Liquid AI, released May 2026. Its published size is 8.47B parameters (1.5B active per token) and a 128k-token context window. Licence: lfm1.0.

52.2±1.0#312 of 719 LLMsfrom 29 benchmark results across 2 sources

$ how it compares

The LLMs ranked either side of LFM2.5 8B-A1B, out of 828 with a score, then more from Liquid AI.

#	model	score	where	RAM	released
373	Yi-Large Preview 01.AI	52±1	comparison only		2024-05
374	Qwen2.5-Math 7B Instruct Alibaba	52±2	local	~6GB	2024-09
375	LFM2.5 8B-A1B Liquid AI	52±1	local	~6.5GB	2026-05
376	GPT-4 (0314) OpenAI	52±1	comparison only		2023-03
377	GLM-4 (0520) Z.ai	52±1	comparison only		2024-05
	more from Liquid AI
370	LFM2.5 2.6B Liquid AI	53±1	local	~3.2GB	2026-07
398	LFM2 24B-A2B Liquid AI	51±1	local	~16GB	2026-02
433	LFM2.5 1.2B Thinking Liquid AI	49±2	local	~2.2GB	2026-01

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasks▁▁▃▆█▆▅▅▅▄▃▁ #213 of 300 · 17% of its score · 8 tests

Beats o1 (2024-12-17), o3-mini, DeepSeek R1 (Jan '25) here, though they rank higher overall. Trails SmolLM3 3B, Jamba 1.5 Mini, Granite 4.1 3B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
BFCL V4 (December 2025 archive) +1	49.2% ✎	▄·█·▄▄▄▄·▄·▄ beats 4/8	medium · avg 43%	▇▇ 4.8%	at its level · several runs
BrowseComp-ZH	13.2% ✎	▄▄·▄▄▄·█···█ beats 4/8	very hard · avg 13%	▇ 2.5%	typical
SWE-bench Verified	0.4% ✎	▄··▄▄█··▄·▄█ beats 0/8	very hard · avg 12%	▇ 2.4%	sharp · noisy
τ²-bench Retail lab	39.8% ✎	▁····▃▁▃▆█▃▁ beats 3/44	hard · avg 31%	▇ 2.0%	at its level · small run
SWE-bench Pro (split unspecified)	0.4% ✎	█▂···▂···▄·▂ beats 0/8	very hard · avg 4.3%	▇ 1.6%	sharp · too hard for it · noisy
τ³-bench Banking	3.4% ✎	▂▂·▅█▂▂▃▂▅▂▃ beats 2/20	very hard · avg 3.7%	▇ 1.4%	blunt · too hard for it
τ²-bench Telecom lab	88.1% ✎	▄█▄········█ beats 5/5	hard · avg 24%	▇ 1.3%	sharp · at its level · noisy
BFCL v3 lab	64.4% ✎	▁▁▁▁▁▃▅█▆▅▄▁ beats 120/170	medium · avg 57%	▇ 0.9%	blunt · at its level · small run

coding▁▁▁▃▆██▇▆▄▃▁ #381 of 458 · 18% of its score · 5 tests

Trails Qwen2.5 1.5B Instruct, WizardLM 13B V1.2, Falcon3 3B Instruct here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
LiveCodeBench v6 (window unspecified)	39.8% ✎	▃····█·▃▅▃·▃ beats 1/8	medium · avg 39%	▇▇ 6.2%	at its level
LiveCodeBench Pro 2025Q2 (Easy)	27.8% ✎	▄···█▄··▄█▄▄ beats 2/8	hard · avg 33%	▇▇ 5.0%	at its level
LiveCodeBench Pro 2025Q2 (Medium)	0.0% ✎	█···▂··▄▂··▂ beats 0/8	very hard · avg 2.0%	▇ 2.9%	sharp · too hard for it · noisy
SWE-bench Verified	0.4% ✎	▄··▄▄█··▄·▄█ beats 0/8	very hard · avg 12%	▇ 2.4%	sharp · noisy
SWE-bench Pro (split unspecified)	0.4% ✎	█▂···▂···▄·▂ beats 0/8	very hard · avg 4.3%	▇ 1.6%	sharp · too hard for it · noisy

chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #126 of 393 * · 1.7% of its score · 1 test

Beats LongCat 2.0 (1.6T-A48B), o3-mini, K-EXAONE (236B-A23B) here, though they rank higher overall.

test	result	vs the others in its run	difficulty	counts	why
IFEval (aggregation unspecified) +1	90.8% ✎	▄▄····▄··███ beats 6/8	easy · avg 82%	▇ 1.7%	small run · several runs

math▁▂▃▅▆▅▄█▄▁▁▁ #212 of 398 · 42% of its score · 4 tests

Beats DeepSeek V3 (Dec '24), GPT-4o (2024-08-06), Command A (03-2025) here, though they rank higher overall. Trails Olmo 3 7B Instruct, DeepSeek R1 Distill Llama 8B, Phi-4-mini-reasoning 3.8B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
AIME 2025 +1	46.0% ✎	█▄▄▄▄····█·▄ beats 4/8	medium · avg 42%	▇▇▇▇▇▇ 15%	sharp · at its level · several runs
HMMT February 2026	38.5% ✎	▅··▃▃·▃··▃·█ beats 4/8	hard · avg 30%	▇▇▇▇ 11%	sharp · at its level
AIME 2026 +1	56.7% ✎	▄·▄█·▄▄···█▄ beats 4/8	medium · avg 48%	▇▇▇ 8.3%	sharp · at its level
MATH-500 +1	93.2% ✎	██▄▄▄··▄···▄ beats 5/8	very easy · avg 89%	▇▇▇ 8.1%	several runs

web search█·█▄██·▄▄·▄▄ #9 of 13 * · 2.5% of its score · 1 test

Beats Nemotron 3 Nano 4B, Gemma 4 E4B, LFM2.5 2.6B here, though they rank higher overall. Trails Qwen3.5 2B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
BrowseComp-ZH	13.2% ✎	▄▄·▄▄▄·█···█ beats 4/8	very hard · avg 13%	▇ 2.5%	typical

other tests7 not in a task · 25% of its score

test	result	vs the others in its run	difficulty	counts	why
MMLU-Pro	63.1% ✎	▂··▃█·▂▂···▂ beats 1/8	medium · avg 59%	▇▇ 5.1%	at its level
MMLU-Redux	80.0% ✎	▂···█··▂▂··▂ beats 3/8	easy · avg 77%	▇▇ 4.7%	at its level
IFBench (aggregation unspecified) +1	51.0% ✎	▅····▃▃·█·▃▃ beats 3/8	medium · avg 43%	▇▇ 3.8%	at its level · several runs
OJBench	8.2% ✎	▃····▃▅··█▃▃ beats 1/8	very hard · avg 12%	▇ 3.1%	typical
SuperGPQA	34.5% ✎	▃·▃·▃▅█····▃ beats 2/8	hard · avg 32%	▇ 2.9%	at its level
Multi-IF (aggregation unspecified) +1	71.4% ✎	▄▄···▄·▄·▄██ beats 4/8	medium · avg 60%	▇ 2.8%	blunt · at its level · several runs
LongBench v2	30.4% ✎	▃······▅▅▃·█ beats 2/8	very hard · avg 32%	▇ 2.4%	typical

$ where to run it

llama.cpp · recommended

version

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 LFM2.5 8B-A1B · needs ~5.2 GB · from huggingface.co/LiquidAI (official)
chat in the terminal
llama cli -hf LiquidAI/LFM2.5-8B-A1B-GGUF:Q4_K_M
Or start a local server and open the address it prints (http://localhost:9931)
llama serve -hf LiquidAI/LFM2.5-8B-A1B-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 LFM2.5 8B-A1B · needs ~5.2 GB · from huggingface.co/LiquidAI (official)
open in LM Studio ↗
Or download it in the terminal
lms get https://huggingface.co/LiquidAI/LFM2.5-8B-A1B-GGUF@Q4_K_M -y
Then run lms ls to see its name, and lms chat followed by that name

Ollama

version

Install Ollama
download it (macOS / Windows) ↗
Linux
curl -fsSL https://ollama.com/install.sh | sh
Run LFM2.5 8B-A1B · needs ~5.58 GB · from ollama.com
chat in the terminal
ollama run lfm2.5:8b

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 LFM2.5 8B-A1B · needs ~4.8 GB, Mac only · from huggingface.co/LiquidAI (official)
chat in the terminal
mlx_lm.chat --model LiquidAI/LFM2.5-8B-A1B-MLX-4bit
Or start an OpenAI-compatible server at http://localhost:8080/v1
mlx_lm.server --model LiquidAI/LFM2.5-8B-A1B-MLX-4bit

Official weights and documentation: huggingface.co/LiquidAI ↗

$ how the score was worked out · openbmb/MiniCPM5-2B model card 81%, LiquidAI/LFM2.5-8B-A1B model… 19%

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
openbmb/MiniCPM5-2B model card	20	▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 81%	Apache-2.0
LiquidAI/LFM2.5-8B-A1B model card · lab-reported	9	▇▇▇▇▇▇ 19%	LFM Open License v1.0
prior · pull toward the middle		▇ 0%
weighted average		= 52.2

not counted (1)

LiquidAI/LFM2.5-8B-A1B model card

AIME 2026 [correct]

50.0%

the only model in this source's run of the test, so there is nothing to compare it with there; another run of the test counts

$ published speed measurements · 3

Reference results for the hardware, engine and settings below. They aren't verified speeds of the downloads above or predictions for your machine.

AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU; Beelink GTR9 Pro 128GB	Q4_K_M	171.17 tok/s	llama.cpp ac4cddeb0 (build_number 9592 as recorded), Vulkan (RADV) · llama-bench; tg128; 128 generated tokens; context depth 0; n_prompt=0 row; n_batch=2048; n_ubatch=512; flash_attn=1; type_k=f16; type_v=f16; n_gpu_layers=999; threads=16; one sequence; file label r10 repetitions Author's own llama-bench CSV row (avg_ts mean over repetitions, stddev_ts retained). GPU string as recorded: Radeon 8060S Graphics (RADV STRIX_HALO). Unified memory; the guide reports ~124GiB OS-visible. Kernel/Mesa/power profile per the run folder notes. Not prompt processing or multi-request throughput. Curated in the guide's data/benchmarks.csv line 30 (2026-06-11, measured-local): Latest upstream Vulkan/RADV control for the current small-MoE speed route; remains a small active-parameter MoE speed result, not a 30B-class capability replacement.	source ↗
AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU; Beelink GTR9 Pro 128GB	Q4_K_M	168.96 tok/s	llama.cpp 2016bf2 (build_number 1 as recorded), Vulkan (RADV) · llama-bench; tg128; 128 generated tokens; context depth 0; n_prompt=0 row; n_batch=2048; n_ubatch=512; flash_attn=-1; type_k=f16; type_v=f16; n_gpu_layers=-1; threads=16; one sequence; file label r5 repetitions Author's own llama-bench CSV row (avg_ts mean over repetitions, stddev_ts retained). GPU string as recorded: Radeon 8060S Graphics (RADV STRIX_HALO). Unified memory; the guide reports ~124GiB OS-visible. Kernel/Mesa/power profile per the run folder notes. Not prompt processing or multi-request throughput. Curated in the guide's data/benchmarks.csv line 35 (2026-06-05, measured-local): Latest/int-dot practical workstation scout; generation-only p0/n128 r20 measured 170.02 t/s. Fastest current small-MoE scout here; not a 30B-class capability replacement.	source ↗
AMD Ryzen AI Max+ 395 (Strix Halo), Radeon 8060S iGPU; Beelink GTR9 Pro 128GB	Q4_K_M	176.48 tok/s	llama.cpp 98d5e8ba8 (build_number 9544 as recorded), Vulkan (RADV) · llama-bench; tg128; 128 generated tokens; context depth 0; n_prompt=0 row; n_batch=2048; n_ubatch=512; flash_attn=1; type_k=f16; type_v=f16; n_gpu_layers=999; threads=16; one sequence; file label r10 repetitions Author's own llama-bench CSV row (avg_ts mean over repetitions, stddev_ts retained). GPU string as recorded: Radeon 8060S Graphics (RADV STRIX_HALO). Unified memory; the guide reports ~124GiB OS-visible. Kernel/Mesa/power profile per the run folder notes. Not prompt processing or multi-request throughput. Curated in the guide's data/benchmarks.csv line 101 (2026-06-07, measured-local): Latest llama.cpp b9544 Vulkan/RADV small-MoE speed control with explicit -dev Vulkan0; r10; latest build did not regress the LFM2.5 speed route.	source ↗

$ 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