5070 Ti owner here (16GB version), running LLMs 24/7 for months. Some real-world notes:

Everyone's recommending models — I'll focus on the stuff nobody tells you about the 5070 Ti specifically:

Model picks for 12GB: +1 for qwen2.5-coder:14b for Python (Q4_K_M fits). For reasoning, qwen3.5:9b over Hermes —

massive quality jump. I'd also try gemma4:e4b as others mentioned, but heads up: it requires think: true in the API or

you get empty responses, and set num_predict: 2048+ because the thinking tokens eat your budget.

What nobody mentions about the 5070 Ti:

- Cap the power (nvidia-smi -pl 200) — sustained LLM inference pushes these cards hard. Mine was crashing with TDR

errors before I capped it at 250W

- Ollama keeps models in VRAM for 5 min after last use. With 12GB, switching between two models = OOM crash. Use

keep_alive: "30s" in your API calls

- num_ctx: 32768 via a custom Modelfile — the default 4K context is useless for real code work

- Skip offloading to RAM with 64GB. A 9B model fully in VRAM at 80 tok/s beats a 30B model half-offloaded at 8 tok/s

for almost everything

LM Studio is fine for testing but switch to Ollama for anything serious — ollama ps, API access, and Modelfiles give

you control you'll need.