What is the best local LLM for an RTX 4070?

On the 12 GB RTX 4070, Qwen 3.5 9B at Q6_K is the best all-round pick — about 9 GB and ~30 tokens/sec. For general chat, Llama 3.1 8B at Q8 is a good alternative. Keep 20B models for 16 GB+ cards.

Can the RTX 4070 run gpt-oss 20B?

Technically yes at Q4_K_M (~12 GB), but it fills the card and leaves almost no room for context, so it is unreliable for real OpenClaw agent loops. A 16 GB card runs gpt-oss 20B comfortably; on a 4070, stick to 9B models or route agents to the cloud.

Is the RTX 4070 or 4070 Ti Super better for AI?

The 4070 Ti Super, clearly — its 16 GB (vs the 4070's 12 GB) is the threshold that lets you run 20B agent models with context. If local OpenClaw tool-calling matters, the Ti Super is worth it.

How much VRAM do I need for local LLMs?

12 GB runs 8-9B assistants well; 16 GB runs 20B agent models; 24 GB runs 27B-class models at a good quant; 32 GB+ adds context headroom and larger MoE models. Match VRAM to the largest model you actually need.

← All guides

Hardware July 6, 2026

Best Local LLM for RTX 4070 (2026): 12GB VRAM Picks

The RTX 4070 (non-Ti) ships 12 GB of VRAM. That makes it a capable small-model card — excellent for 8-9B assistants — but 12 GB is the line below which 20B agent models stop fitting comfortably.

Picking hardware for an OpenClaw host?

Use the local model calculator first, then see our AI training options if you want help matching your workload to the right rig.

Short answer: on a 12 GB RTX 4070, run Qwen 3.5 9B at Q6_K (~9 GB, ~30 tok/sec) as your daily assistant, or Llama 3.1 8B at Q8 for general chat. gpt-oss 20B technically loads at Q4 but leaves almost no room for context — for reliable OpenClaw agent work, either use a 16 GB+ card or route agents to the cloud.

The VRAM Math

The 4070 is a great value chat card but a marginal agent card. If OpenClaw tool-calling reliability matters, the extra 4 GB of a 16 GB card (4070 Ti Super, 4080) is the single most useful upgrade.

What Actually Fits (Model Picks)

Model	Quant	VRAM used	Speed	Notes
Qwen 3.5 9B	Q6_K	~9 GB	~30 tok/s	Best daily assistant
Llama 3.1 8B	Q8_0	~9 GB	~32 tok/s	General chat
gpt-oss 20B	Q4_K_M	~12 GB (very tight)	~18 tok/s	Loads but no context headroom
Qwen 3.5 9B	Q4_K_M	~6 GB	~35 tok/s	Leaves room for bigger context

What You Can’t Run

gpt-oss 20B with usable context — at 12 GB the KV cache pushes you out of memory; keep 20B for 16 GB+ cards.
Qwen 3.6 27B at any good quant — needs 17-18 GB, far past 12 GB.
70B models — not remotely close on a 12 GB card.

🎮 12 GB VALUE — OR 16-24 GB FOR AGENTS

For a linkable 12 GB card the RTX 3060 is the budget option. If you want to run 20B agent models reliably, step to a 16 GB 4070 Ti Super or a 24 GB 4090.

3060MSI RTX 3060 12 GB ↗ 4070TiSGIGABYTE RTX 4070 Ti Super 16 GB ↗ 4090GIGABYTE RTX 4090 24 GB ↗

OpenClaw Setup

Point OpenClaw at your local model through Ollama:

# pull and run your pick, then set it as the OpenClaw default
ollama pull qwen3:9b
openclaw config set agents.defaults.models.chat "ollama/qwen3:9b"

For agent reliability, prefer a model with clean tool-call output (gpt-oss 20B where it fits) and cap context to what your memory holds. See the tool-calling reliability guide.