BestLLMfor EN Your hardware. Your LLM. Your call.
APIOpen data Find my LLM
Editorial ranking · 2026

Best local LLM for 8GB VRAM

Top 8 open-source picks for 8 GB VRAM budgets, ranked by benchmark performance and real-world fit. Updated monthly.

#1

Granite 4.0 H-Tiny 7B-A1B

7B · IBM · Apache 2.0

IBM's edge-class hybrid MoE with 7B total and only 1B active parameters — Apache 2.0 licensed and built for embedded and low-cost serving.

VRAM Q4: 4 GB · Context: 125k
Read full fiche →
#2

Mistral Nemo 12B Instruct

12B · Mistral AI · Apache 2.0

Mistral AI and NVIDIA's co-developed 12B instruct model with 128k context, the Tekken tokenizer, and strong European multilingual coverage.

VRAM Q4: 7 GB · Context: 125k
Read full fiche →
#3

Gemma 3 12B

12B · Google · Gemma

The 12B sweet spot of Google's Gemma 3 line — multimodal, 128K context, and 140 languages. Fits on a single consumer GPU with room for batching.

VRAM Q4: 7 GB · Context: 125k
Read full fiche →
#4

Nemotron Nano v2 VL 12B

12.6B · NVIDIA · NVIDIA Open Model License

NVIDIA's 12.6B enterprise VLM with strong DocVQA and ChartQA scores, tuned for professional document extraction workflows.

VRAM Q4: 8 GB · Context: 125k
Read full fiche →
#5

Lucie 7B

7B · OpenLLM-France · Apache 2.0

A French-sovereign 7B model from OpenLLM-France, backed by CNRS and LINAGORA, with a fully transparent and auditable training corpus.

VRAM Q4: 5 GB · Context: 4k
Read full fiche →
#6

DeepSeek R1 Distill 7B

7B · DeepSeek · MIT

A 7B DeepSeek model distilled from R1 671B with explicit chain-of-thought reasoning. Surprisingly strong on AIME and MATH for its size.

VRAM Q4: 5 GB · Context: 32k
Read full fiche →
#7

Qwen 3 8B

8B · Alibaba · Apache 2.0

Alibaba's 8B dense model with a toggleable thinking mode and broad multilingual coverage. Punches well above its weight for an 8B and runs comfortably on a single consumer GPU.

VRAM Q4: 5 GB · Context: 128k
Read full fiche →
#8

Qwen 2.5 VL 7B

7B · Alibaba · Apache 2.0

A 7B vision-language model from Alibaba with state-of-the-art results in its class, scoring 95.7 on DocVQA. Handles hour-long video, bounding-box grounding, and multilingual OCR.

VRAM Q4: 6 GB · Context: 125k
Read full fiche →