What fits in a RTX 3070's 8 GB?

Ampere. 8 GB VRAM, verified from the vendor datasheet. Cheapest live rate from $0.050/hr across 2 providers.

Models that fit at FP16 (8 GB)

• Qwen/Qwen3-0.6B (752M)
• openai-community/gpt2 (137M)
• Qwen/Qwen2.5-1.5B-Instruct (1.5B)
• Qwen/Qwen2.5-3B-Instruct (3.1B)
• openai/whisper-large-v3-turbo (809M)
• Qwen/Qwen3-Embedding-0.6B (596M)
• meta-llama/Llama-3.2-1B-Instruct (1.2B)
• Qwen/Qwen2.5-0.5B-Instruct (494M)
• openai/whisper-large-v3 (1.5B)
• Qwen/Qwen3-ASR-1.7B (2.3B)
• Qwen/Qwen3-1.7B (2.0B)
• EleutherAI/pythia-160m (213M)
• google/gemma-3-1b-it (1000M)
• baidu/Unlimited-OCR (3.3B)
• RadixArk/Kimi-K3-DSpark (2.2B)
• openai/whisper-small (242M)
• Qwen/Qwen3-VL-2B-Instruct (2.1B)
• Qwen/Qwen3.5-2B (2.3B)
• microsoft/Florence-2-base (232M)
• HuggingFaceTB/SmolLM2-135M (135M)
• MahmoudAshraf/mms-300m-1130-forced-aligner (315M)
• deepseek-ai/DeepSeek-OCR (3.3B)
• Qwen/Qwen3.5-0.8B (873M)
• zai-org/GLM-OCR (1.3B)
• Qwen/Qwen3-VL-Reranker-2B (2.1B)
• mlx-community/parakeet-tdt-0.6b-v2 (618M)
• prism-ml/Bonsai-27B-mlx-1bit (1.7B)
• Qwen/Qwen3-1.7B-Base (1.7B)
• Qwen/Qwen2-VL-2B-Instruct (2.2B)
• Qwen/Qwen2.5-0.5B (494M)
• mlx-community/parakeet-tdt-0.6b-v3 (627M)
• stabilityai/stable-diffusion-xl-base-1.0 (2.6B)
• vikhyatk/moondream2 (1.9B)
• gigant/romanian-wav2vec2 (315M)
• comodoro/wav2vec2-xls-r-300m-cs-250 (315M)
• nvidia/parakeet-ctc-1.1b (1.1B)
• stable-diffusion-v1-5/stable-diffusion-v1-5 (860M)
• HuggingFaceTB/SmolVLM2-500M-Video-Instruct (507M)
• microsoft/phi-2 (2.8B)
• KBLab/wav2vec2-large-voxrex-swedish (315M)

Models that fit at INT4 (quantized, 8 GB)

• Qwen/Qwen3-0.6B (752M)
• openai-community/gpt2 (137M)
• Qwen/Qwen3-8B (8.2B)
• Qwen/Qwen3.5-9B (9.7B)
• Qwen/Qwen2.5-7B-Instruct (7.6B)
• Qwen/Qwen3-VL-8B-Instruct (8.8B)
• Qwen/Qwen2.5-VL-7B-Instruct (8.3B)
• Qwen/Qwen2.5-1.5B-Instruct (1.5B)
• Qwen/Qwen2.5-3B-Instruct (3.1B)
• Qwen/Qwen3.5-4B (4.7B)
• openai/whisper-large-v3-turbo (809M)
• Qwen/Qwen3-Embedding-0.6B (596M)
• meta-llama/Llama-3.2-1B-Instruct (1.2B)
• Qwen/Qwen3-4B (4.0B)
• Qwen/Qwen2.5-0.5B-Instruct (494M)
• meta-llama/Llama-3.1-8B-Instruct (8.0B)
• openai/whisper-large-v3 (1.5B)
• google/gemma-4-E4B-it (8.0B)
• Qwen/Qwen3-ASR-1.7B (2.3B)
• Qwen/Qwen3-VL-4B-Instruct (4.4B)
• Qwen/Qwen2.5-VL-3B-Instruct (3.8B)
• Qwen/Qwen3-1.7B (2.0B)
• EleutherAI/pythia-160m (213M)
• Qwen/Qwen3-4B-Instruct-2507 (4.0B)
• google/gemma-4-E2B-it (5.1B)
• google/gemma-4-12B-it (12.0B)
• Qwen/Qwen3-Embedding-4B (4.0B)
• google/gemma-3-1b-it (1000M)
• baidu/Unlimited-OCR (3.3B)
• RadixArk/Kimi-K3-DSpark (2.2B)
• openai/whisper-small (242M)
• datalab-to/chandra-ocr-2 (5.3B)
• Qwen/Qwen3-VL-2B-Instruct (2.1B)
• Qwen/Qwen3.5-2B (2.3B)
• microsoft/Florence-2-base (232M)
• HuggingFaceTB/SmolLM2-135M (135M)
• MahmoudAshraf/mms-300m-1130-forced-aligner (315M)
• deepseek-ai/DeepSeek-OCR (3.3B)
• Qwen/Qwen3.5-0.8B (873M)
• Qwen/Qwen2.5-Coder-7B-Instruct (7.6B)

Caveat: Requires a quantized checkpoint actually published for the model, check its Hugging Face page before relying on this.