Best LLMs for 4 GB
Pick your memory. We rank every catalogued model that fits at its recommended quantization, by task, and show what the next tier up unlocks.
Written by Jakub Rusinowski · Last updated
Memory tier4 GBFor example RTX 5050, RTX 5060, RTX 4060
Models that fit32entirely in memory
Best at 4 GB, by task
Score is the task-specific quality score on a 0–100 scale; speed is estimated for the RTX 5050.
The strongest that fit
Top 15, by logic and coding
| Model | Memory · marker = 4 GB | VRAM | Speed on RTX 5050 (est.) | Score |
|---|---|---|---|---|
| Gemma 3 4B Instruct Gemma 3 · Q4_K_M | 2.4 GB | ~83 tok/s | 80 | |
| Phi-4 Mini (3.8B) Phi-4 Mini · Q4_K_M | 2.3 GB | ~79 tok/s | 79 | |
| Qwen 3.5 4B Qwen 3.5 · Q4_K_M | 2.4 GB | ~86 tok/s | 79 | |
| VibeThinker 3B VibeThinker · Q4_K_M | 1.9 GB | ~94 tok/s | 79 | |
| Phi 3.5 Mini Phi 3.5 Family · Q4_K_M | 2.3 GB | ~82 tok/s | 79 | |
| BitNet b1.58 3B BitNet b1.58 · 1.58-bit | 0.7 GB | ~160 tok/s | 74 | |
| StarCoder 2 3B StarCoder 2 · Q4_K_M | 1.8 GB | ~107 tok/s | 74 | |
| EXAONE 3.5 2.4B EXAONE 3.5 · Q4_K_M | 1.4 GB | ~115 tok/s | 73 | |
| Aya 3B Aya 3B (Tiny Aya) · Q4_K_M | 1.8 GB | ~96 tok/s | 72 | |
| Qwen 3.5 2B Qwen 3.5 · Q4_K_M | 1.2 GB | ~124 tok/s | 72 | |
| Granite 4.1 3B IBM Granite 4.1 · Q4_K_M | 1.8 GB | ~96 tok/s | 71 | |
| Ministral 3B Ministral · Q4_K_M | 2.2 GB | ~84 tok/s | 71 | |
| Falcon 3 3B Instruct Falcon 3 · Q4_K_M | 2 GB | ~93 tok/s | 69 | |
| SmolLM3 3B SmolLM3 · Q4_K_M | 1.9 GB | ~94 tok/s | 69 | |
| VibeThinker 1.5B VibeThinker · Q4_K_M | 0.9 GB | ~145 tok/s | 69 |
What 6 GB unlocks
The strongest models that need the next tier up.
FAQ
What LLMs run well with 4 GB VRAM?
Gemma 3, Phi-4 Mini, Qwen 3.5, VibeThinker, Phi 3.5 Family all fit in 4 GB VRAM.
Which GPUs have 4 GB VRAM?
NVIDIA GeForce RTX 5050, NVIDIA GeForce RTX 5060, NVIDIA GeForce RTX 4060, AMD Radeon RX 9060 XT 8GB.
All tiers at a glance
| Memory | Models fit | Best overall | Largest that fits |
|---|---|---|---|
| 4 GB | 32 | Gemma 3 4B Instruct | Gemma 3 4B Instruct · 2.4 GB |
| 6 GB | 56 | Granite 3.0 8B Instruct | Qwen 2.5 VL 7B Instruct · 5 GB |
| 8 GB | 71 | Granite 3.0 8B Instruct | Falcon 3 10B Instruct · 6.2 GB |
| 10 GB | 83 | Qwen 2.5 14B Instruct | Qwen 3 14B · 8.9 GB |
| 12 GB | 85 | Qwen 2.5 14B Instruct | StarCoder 2 15B · 9.4 GB |
| 16 GB | 95 | Qwen 2.5 14B Instruct | Mistral Small 3 (24B) · 14.5 GB |
| 20 GB | 111 | Qwen 2.5 14B Instruct | Gemma 4 31B · 18.7 GB |
| 24 GB | 131 | Qwen 2.5 Coder 32B | Qwen 3.6 35B-A3B · 21.1 GB |
| 32 GB | 132 | Qwen 2.5 Coder 32B | Qwen 3.6 35B-A3B · 21.1 GB |
| 48 GB | 138 | Qwen 2.5 72B Instruct | Qwen 2.5 VL 72B Instruct · 44.3 GB |
| 64 GB | 143 | Qwen 2.5 72B Instruct | Qwen3-Coder-Next (80B-A3B MoE) · 48.3 GB |
| 80 GB | 152 | Qwen 2.5 72B Instruct | Qwen 3.5 122B-A10B · 73.7 GB |
Buy this hardware Intel Arc B570 10GB — 10 GB VRAM · 150 W board powerDeploy in the cloud now RTX 4090 on RunPod — from $0.34/hr · rate checked 2026-07
or compare on Vast.ai from $0.35/hr (typical low · varies)
As an Amazon Associate we earn from qualifying purchases. Cloud GPU links are referral links — we may earn a commission at no extra cost to you.
Can-I-Run Pages Near 4 GB
- BitNet b1.58 on NVIDIA GeForce RTX 5060 Ti 8GB
- BitNet b1.58 on NVIDIA GeForce RTX 5060
- BitNet b1.58 on NVIDIA GeForce RTX 4060
- DeepSeek R1 on NVIDIA GeForce RTX 5060 Ti 8GB
- DeepSeek R1 on NVIDIA GeForce RTX 5060
- DeepSeek R1 on NVIDIA GeForce RTX 4060
- DeepSeek R1 on NVIDIA GeForce RTX 3070 Ti
- DeepSeek R1 on NVIDIA GeForce RTX 3070