Molmo2 4B — Hardware Requirements & GPU Compatibility
Vision
Molmo2-4B is Allen Institute for AI's (Ai2) vision-language model for image, video, and multi-image understanding and grounding, built on a Qwen3-4B-Instruct backbone with a SigLIP 2 vision encoder. It is trained on Ai2's own curated Molmo2 datasets rather than third-party captioning data of unclear provenance, and beyond ordinary visual question answering it supports pointing at and tracking objects across video frames. The card reports state-of-the-art results among open weight-and-data models on short-video understanding, counting, and captioning, with competitive results on long videos. At under 5 billion parameters, it fits on a single consumer GPU. Context length is 36,864 tokens. It is released under the Apache 2.0 license, permitting unrestricted commercial and research use, and was published in December 2025.
Runs great on a 24 GB GPU
Runs from 8 GB (Q4_K_M · 36K)Best on 24 GB (Q4_K_M · ~193 tok/s · 36K)Macs from 8 GB (Q4_K_M · ~21 tok/s)
Apache 2.0
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run Molmo2 4B
One recommended way. Every other method, OS and quant is in the step-by-step guide.Each way opens a step-by-step guide for your OS and quant.
Recommended
ollama run hf.co/mu2solutions/Molmo2-4B-GGUF:Q4_K_MQ4_K_M fits fully on the RTX 4090 · 19.3 GB free · 16K context
Download: Molmo2-4B-text-Q4_K_M.gguf2.72 GBmu2solutions
Ways to runOther ways
each opens step-by-stepRun in cloud
Runs on RTX 3060 12GB (7 GB headroom) · Q4_K_M
- Generation speed
- ~69 tok/s
- generation speed
- Cost per 1M output tokens
- $0.24
- per 1M output tokens
How Much VRAM Does Molmo2 4B Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | On RTX 4090 (reference) | |
|---|---|---|---|---|
| 2.5 GB | 5.4 GB | 2-bit · K-quant — 2-bit quantization with K-quant improvements | Fits: 36K context · ~257 tok/s | |
| 2.9 GB | 5.7 GB | 3-bit medium — 3-bit medium quantization | Fits: 36K context · ~230 tok/s | |
| 3.4 GB | 6.2 GB | 4-bit medium — 4-bit medium quantization — most popular sweet spot | Fits: 36K context · ~193 tok/s | |
| 3.9 GB | 6.8 GB | 5-bit medium — 5-bit medium quantization — good quality/size tradeoff | Fits: 36K context · ~166 tok/s | |
| 4.5 GB | 7.3 GB | 6-bit, very good — 6-bit quantization, very good quality | Fits: 36K context · ~146 tok/s | |
| 5.3 GB | 8.2 GB | 8-bit, near-lossless — 8-bit quantization, near-lossless | Fits: 36K context · ~123 tok/s | |
| 10.2 GB | 13 GB | 16-bit · bfloat — Brain floating point 16 — preferred for training | Fits: 36K context · ~64 tok/s |
Selected Q4_K_M
4-bit medium quantization4.7GBat 16K context
- Weights
- 2.9 GB
- KV cache
- 1.5 GB
- Overhead
- 0.3 GB
Fits RTX 4090 (reference) · 19.3 GB free · ~193 tok/s
≥50 tok/s needs ≥261 GB/s · RTX 5090 ~343, RTX 3090 Ti ~193
Which Hardware Can Run Molmo2 4B?
Q4_K_M · 3.4 GBMolmo2 4B (Q4_K_M) requires 3.4 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 5+ GB is recommended. Using the full 37K context window can add up to 3.2 GB, bringing total usage to 6.6 GB. 52 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti. 163 devices with unified memory can run Molmo2 4B, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), MacBook Air 13" M3 (8 GB).
≤ 11 GB
≤ 11 GB tier →fits · 3.3 GB free · up to 36K ctx
16 GPUs · 1 Mac class · 2 device families
RTX 3080~145Simulate output on RTX 3080RTX 2080 Ti~118Simulate output on RTX 2080 TiRTX 3070 Ti~116Simulate output on RTX 3070 TiM3*8 GB~21Simulate output on iMac 24" M3 (8 GB)Orin Nano~20Simulate output on NVIDIA Jetson Orin Nano 8GB (Super)iPhone~14Simulate output on iPhone 1720 GB
fits · 15.3 GB free · up to 36K ctx
1 GPU
RX 7900 XT~141Simulate output on RX 7900 XT32–40 GB
fits · 27.3 GB free · up to 36K ctx
1 GPU · 3 Mac classes · 4 device families
RTX 5090~343Simulate output on RTX 5090M5 Max36 GB~95Simulate output on Mac Studio M5 Max (36 GB)M532 GB~32Simulate output on MacBook Air 13" M5 (32 GB)Orin~39Simulate output on NVIDIA Jetson AGX Orin 32GBSnapdragon X Elite (X1E-80-100)~26Simulate output on Snapdragon X Elite Copilot+ PC
By VRAM tier · typical GPUs and Macs at Q4_K_M
Alternatives to Molmo2 4B
OLMo family · 20 models →Similar models
6| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| Qwen3.5 4BAlibaba | 98.5 | 3.4 GB | ~201 | same size |
| WeVisOne VL 4BTencent | — | 4.7 GB | ~193 | same size |
| FrogNano 4B 2609Microsoft | — | 3.4 GB | ~201 | same size |
| Gemma 4 E2BGoogle | — | 3.5 GB | ~191 | same size |
| Qwen3.6 27B MTPLX Optimized SpeedYoussofal | — | 4 GB | ~185 | same size |
| AFM 4.5BArcee AI | — | 4.3 GB | ~204 | same size |
Specifications
- Publisher
- Allen AI
- Family
- OLMo
- Parameters
- 4.9B
- Architecture
- Molmo2ForConditionalGeneration
- Context Length
- 36,864 tokens
- Vocabulary Size
- 151,936
- Release Date
- 2025-12-14
- License
- Apache 2.0
Frequently Asked Questions
How much VRAM does Molmo2 4B need?
Molmo2 4B requires 3.4 GB of VRAM at Q4_K_M, or 10.2 GB at BF16. Full 37K context adds up to 3.2 GB (6.6 GB total).
VRAM = Weights + KV Cache + Overhead
Weights = 4.9B × 4.8 bits ÷ 8 = 2.9 GB
KV Cache + Overhead ≈ 0.5 GB (at 2K context + ~0.3 GB framework)
Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.
KV Cache + Overhead ≈ 3.7 GB (at full 37K context)
VRAM usage by quantization
What's the best quantization for Molmo2 4B?
For Molmo2 4B, Q4_K_M (3.4 GB) offers the best balance of quality and VRAM usage. Q5_K_M (3.9 GB) provides better quality if you have the VRAM. The smallest option is Q2_K at 2.5 GB.
VRAM requirement by quantization
★ Recommended — best balance of quality and VRAM usage.
Can I run Molmo2 4B on a Mac?
Yes — Mac Mini M2 (8 GB) and 145 other Macs can run Molmo2 4B. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 3.4 GB of usable unified memory (RAM minus macOS overhead).
Can I run Molmo2 4B locally?
Yes — Molmo2 4B can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is Molmo2 4B?
At Q4_K_M, Molmo2 4B can reach ~1412 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~193 tok/s. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.
tok/s = (bandwidth GB/s ÷ model GB) × efficiency
Example: NVIDIA B200 → 8000 ÷ 3.4 × 0.65 = ~1529 tok/s
Estimated speed at Q4_K_M (3.4 GB)
Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
What's the download size of Molmo2 4B?
At Q4_K_M, the download is about 2.91 GB. The full-precision BF16 version is 9.70 GB. The smallest option (Q2_K) is 2.06 GB.
Which GPUs can run Molmo2 4B?
52 consumer GPUs can run Molmo2 4B at Q4_K_M. Top options include AMD Radeon RX 6700 XT, AMD Radeon RX 6800, AMD Radeon RX 6800 XT. 52 GPUs have plenty of headroom for comfortable inference.
Which devices can run Molmo2 4B?
163 devices with unified memory can run Molmo2 4B at Q4_K_M (3.4 GB), including AMD Ryzen AI 9 HX 370 (Strix Point) Laptop, ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.