Molmo 2 8B

+ contribute

Molmo 2 8B is an open-weights language model from Ai2, released December 2025. Its published size is 8.66B parameters. Licence: apache-2.0.

Catalogue identity notes

Comparison-only. Arena entry 'molmo-2-8b' (text, vision). Maker repo allenai/Molmo2-8B @e28fa28: licence tag apache-2.0. 8,661,703,120 parameters from the Hugging Face safetensors metadata (dense, so active = total). Released 2025-12: Ai2 Molmo 2 announcement dated December 11, 2025.

55.2±1.8#262 of 719 LLMsfrom 15 benchmark results across 2 sources

$ how it compares

The LLMs ranked either side of Molmo 2 8B, out of 828 with a score, then more from Ai2.

#	model	score	where	RAM	released
312	Magistral Small 1.0 (2506) Mistral	55±1	local	~16GB	2025-06
313	Qwen3 30B-A3B (Apr 2025, hybrid) Alibaba	55±1	api $0.35/M +local	~19GB	2025-04
314	Molmo 2 8B Ai2	55±2	local	~6.8GB	2025-12
315	Magistral Medium 1.1 (2507) Mistral	55±2	comparison only		2025-07
316	Gemini 1.5 Pro Experimental (0801) Google	55±1	comparison only		2024-08
	more from Ai2
259	Olmo 3.1 32B Think Ai2	58±1	local	~21GB	2025-12
262	Olmo 3 32B Think Ai2	58±1	local	~21GB	2025-11
326	Olmo 3.1 32B Instruct Ai2	55±2	local	~21GB	2025-12

see all LLMs · find the ones that fit your machine

$ what it's good at

chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #195 of 393 * · 28% of its score · 1 test

Beats Granite 4.2 8B, Qwen3-VL 30B-A3B Thinking, Mistral Small 4 (119B-A6.5B) here, though they rank higher overall. Trails LFM2.5 1.2B Instruct, LFM2.5 1.2B Thinking, Gemma 3 12B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
LMArena (text)	1321.8 ✎	▁▁▁▂▃▃▃▆▆▇█▃ beats 155/379	avg 1268	▇▇▇▇▇▇▇▇▇▇▇ 28%	blunt

math▁▂▃▅▆▅▄█▄▁▁▁ #282 of 398 * · 8.6% of its score · 1 test

Trails DeepSeek R1 Distill Qwen 1.5B, Phi-2 (2.7B), Granite 8B Code Base (4K) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
MathVista (testmini) lab	58.9% ✎	▁··▁▄▄▇█▅▆█▄ beats 14/47	medium · avg 60%	▇▇▇ 8.6%	at its level · small run

vision & documents▁▁▃▂▁▃▄▂▆▇█▄ #109 of 147 · 70% of its score · 14 tests

Beats DeepSeek V4 Flash Vision (Exp), GPT-4 Turbo (2024-04-09) here, though they rank higher overall. Trails Gemma 3 27B, Mistral Small 3.1 24B (2503), Nova Premier here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Arena vision (style controlled)	1111.7 ✎	▁▁▂▂▂▂▄▅▅▆██ beats 31/145	avg 1068	▇▇▇▇▇▇▇▇ 20%	blunt
MMMU (val) lab	53.0% ✎	▁····▁▁▄█▇▆▆ beats 14/52	medium · avg 54%	▇▇▇▇ 10%	at its level · small run
MathVista (testmini) lab	58.9% ✎	▁··▁▄▄▇█▅▆█▄ beats 14/47	medium · avg 60%	▇▇▇ 8.6%	at its level · small run
InfographicVQA lab	80.1% ✎	▃·▃▅█▃▃▃▃▃▃▃ beats 8/13	easy · avg 79%	▇▇▇ 6.9%	small run
CountBench lab	93.7% ✎	▂▂▃▂·▂▂▃▃·▅█ beats 9/18	very easy · avg 91%	▇▇ 5.5%	maxed out · small run
VQAv2 (val) lab	87.0% ✎	▃▅▃█·▃▅·▃·▃▅ beats 13/13	easy · avg 81%	▇ 3.5%	blunt · small run
ChartQA lab	86.0% ✎	▁··▁▁▁▁▁▂▅▅█ beats 26/44	easy · avg 83%	▇ 2.8%	blunt · small run
DocVQA lab	93.2% ✎	▁··▃▁▁▁▇█▇▄▅ beats 8/28	very easy · avg 93%	▇ 2.7%	blunt · maxed out · small run

6 more tests

test	result	vs the others in its run	difficulty	counts	why
RealWorldQA lab	77.6% ✎	▂▃▂█▃██▆█▂▆▆ beats 23/38	easy · avg 73%	▇ 2.6%	blunt · small run
Molmo2-VideoCount accuracy (val) lab	35.5% ✎	▃··▃·▃····█▃ beats 3/6	hard · avg 22%	▇ 2.4%	blunt · at its level
Ai2 Molmo2 human preference Elo (video/image QA, 2026) lab	1057.0 ✎	▄··▄··▄··▄█▄ beats 2/6	avg 975	▇ 2.1%	blunt
AI2D test lab	95.8% ✎	▁▁▅▇▃▇·▄·█·▃ beats 27/28	easy · avg 86%	▇ 1.3%	blunt · small run
TextVQA lab	85.7% ✎	▁▁▁▄▅▁▃·▃▃▃█ beats 24/24	easy · avg 79%	▇ 1.3%	blunt · small run
Molmo2-CapTest dense video captioning F1 lab	43.2% ✎	█···█·████·█ beats 3/6	hard · avg 32%	▇ 1.1%	blunt · at its level

$ 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 Molmo 2 8B · needs ~5 GB · from huggingface.co/mu2solutions
chat in the terminal
llama cli -hf mu2solutions/Molmo2-8B-GGUF:Q4_K_M
Or start a local server and open the address it prints (http://localhost:9931)
llama serve -hf mu2solutions/Molmo2-8B-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 Molmo 2 8B · needs ~5 GB · from huggingface.co/mu2solutions
open in LM Studio ↗
Or download it in the terminal
lms get https://huggingface.co/mu2solutions/Molmo2-8B-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 Molmo 2 8B · needs ~5 GB · from huggingface.co/mu2solutions
chat in the terminal
ollama run hf.co/mu2solutions/Molmo2-8B-GGUF:Q4_K_M

Other ways to run it

Run it as published · needs ~34.65 GB, NVIDIA only · from github.com/vllm-project
vllm serve allenai/Molmo2-8B --trust-remote-code

Weights and documentation: huggingface.co/allenai ↗

$ how the score was worked out · Molmo2: Open Weights and Data… 50%, LMArena leaderboard dataset 48%

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
Molmo2: Open Weights and Data for Vision-Language Models with Video Understanding and Grounding (Ai2, arXiv 2601.10611v4) · lab-reported	13	▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 50%	CC BY 4.0
LMArena leaderboard dataset	2	▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 48%	CC BY 4.0
prior · pull toward the middle		▇ 1%
weighted average		= 55.2

$ 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