Model

Molmo 2 8B

OPEN

Allen Institute for AI (Ai2) · Molmo 2 family · released Dec 11, 2025

Latest open multimodal flagship (Qwen3-based); extends Molmo's grounded pointing to video with frame-level spatial/temporal grounding and multi-object tracking. Note: weights are Apache 2.0 but some training datasets are restricted to academic/non-commercial research use per Ai2's own blog.

ReasoningCodingVisionFunction callingTool useAgentic

904.2

Elo · rank #234

Parameters

8.66B

Active params

8.66B (dense)

Context

36K tokens

Architecture

Dense transformer (Qwen3-8B LLM decoder + SigLIP2 vision encoder)

License

Apache 2.0

Languages

API price (in/out)

No hosted API

Modalities

text · vision · video

Benchmark results

Bar shows position within the tracked field; marker = field best

AI2DVision95.8%#3

best: Molmo 72B · 96.3%

ARC-ChallengeReasoning89.6%#27

best: Llama 3.1 405B · 96.9%

ChartQAVision86.0%#20

best: MiniMax-VL-01 · 91.7%

DocVQAVision93.2%#20

best: Qwen2-VL-72B · 96.5%

GSM8KMath89.7%#42

best: Llama 3.1 405B · 96.8%

MathVistaVision58.9%#43

best: Seed 2.1 Pro · 90.7%

MBPP+Coding57.5%#31

best: Llama 3.1 405B · 88.6%

MMLUKnowledge76.6%#118

best: OpenAI o3 · 92.9%

MMMUVision53.0%#90

best: Claude Fable 5 · 89.3%

TextVQAVision85.7%#1

best: this model · 85.7%

Video-MMEVision69.9%#18

best: Seed 2.1 Pro · 89.2%

VQAv2Vision87.0%#2

best: Reka Edge (2026) · 88.4%

Run it locally

VRAM @ Q4

5 GB

VRAM @ FP16

18 GB

Fits on (Q4)

RTX 3060 12GBRTX 4070 Ti 16GBRTX 3090 24GBRTX 4090 24GBRTX 5090 32GBM4 Pro 48GBM3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB

No GGUF/llama.cpp support found (custom molmo2 architecture); MLX quants available for Apple Silicon.

Quantizations

MLX

Molmo 2 family

Elo progression across releases

Molmo 2 8BDec 2025904.2

Resources

HFallenai/Molmo2-8B GHallenai/molmo2 DOCSallenai.org/blog/molmo2 APIAvailable via Allen Institute for AI (Ai2) API

API price weights · each benchmark row carries its own source badge (see methodology)