Best LLMs for 8GB VRAM in 2026 — ranked by measured file sizes
Every pick below comfortably fits 8GB VRAM at the listed quant with 8K context. Ranking = largest model (more capable) that still fits, ties broken by real download counts. Sizes are measured GGUF files from Hugging Face, updated daily — not formula estimates.
8GB is the entry point where local LLMs get practical: 7-8B models at Q4-Q6, or 12-14B MoE models (few active params) at comfortable speed. This is the most common laptop-GPU tier.
| Model | Params | Best quant that fits | VRAM needed | Downloads |
|---|---|---|---|---|
| Ternary-Bonsai-2-27B | 27B | BF16 | 2.6 GB | 2336k |
| Humanizer-12B | 12B | IQ3_XXS | 7.3 GB | 4k |
| gemma-4-12b-it | 12B | UD-Q2_K_XL | 7.1 GB | 1499k |
| gemma-3-12b-it-qat | 12B | UD-IQ1_S | 7.3 GB | 93k |
| Qwen3.5-9B-Kimi-k3-Distilled-i1 | 9B | Q3_K_M | 7.2 GB | 13k |
| Meta-Llama-3.1-8B-Instruct | 8B | Q3_K_XL | 7.2 GB | 488k |
| granite-4.2-8b | 8B | Q3_K_L | 7.1 GB | 251k |
| Qwen2.5-VL-7B-Instruct | 8B | Q4_1 | 7.3 GB | 223k |
| Qwen3.8-4B-Distill | 4B | Q8_0 | 7.0 GB | 541k |
| Qwen3-VL-4B-Instruct | 4B | Q8_0 | 6.6 GB | 462k |
| gemma-3n-E4B-it | 4B | Q4_1 | 7.1 GB | 35k |
| Spark-X2.6-4B | 4B | Q8_0 | 6.8 GB | 1k |
Leaves ~8% headroom for fragmentation. Long contexts (>8K) grow the KV cache — open the model page for exact math.
How we pick
A model qualifies when its measured GGUF file plus KV cache plus runtime overhead stays under 8GB x 0.92. We choose the largest such quant (bigger quant = less degradation), then rank models by parameter count — at 8GB VRAM the best model you can run is almost always the largest one that still fits. MoE models count their total size (all experts live in memory), not just active params.