What is the best local LLM for an RTX 4080 or 4080 Super?

gpt-oss 20B at Q4_K_M is the best pick for OpenClaw agent work on a 16 GB RTX 4080 or 4080 Super — it uses about 12-13 GB and produces clean tool-call JSON at roughly 30 tokens/sec. For maximum quality on a smaller model, Qwen 3.5 9B at Q8 fits comfortably.

Can the RTX 4080 run Qwen 3.6 27B?

Only at a degraded quant. Qwen 3.6 27B needs about 17-18 GB at Q4_K_M, which overflows the 4080's 16 GB. It fits at Q3_K_S (~14 GB) but code quality drops noticeably. For 27B at a good quant, step up to a 24 GB card.

Is the RTX 4080 Super better than the 4080 for local LLMs?

Marginally. The 4080 Super has slightly higher memory bandwidth (~736 vs ~717 GB/s), which translates to a few percent more tokens/sec. Both have the same 16 GB VRAM, so they fit exactly the same models — the choice is about price, not capability.

Can an RTX 4080 run a 70B model?

Not usefully. A 70B model needs roughly 40 GB at Q4, far beyond the 4080's 16 GB. You would need a 24 GB+ card at very low quants, two GPUs, or high-unified-memory Apple Silicon for 70B-class local inference.

← All guides

Hardware July 7, 2026

Best Local LLM for RTX 4080 & 4080 Super (2026): 16GB VRAM Picks

The RTX 4080 and 4080 Super both ship 16 GB of VRAM, which puts them squarely in the 'gpt-oss 20B at Q4' tier for local LLMs and OpenClaw agents.

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 16 GB RTX 4080 or 4080 Super, run gpt-oss 20B at Q4_K_M for OpenClaw agent work (clean tool calls, ~30 tok/sec) or Qwen 3.5 9B at Q8 when you want maximum quality on a smaller model. 16 GB is a real 16 GB card — it does not fit Qwen 3.6 27B at a good quant, and 70B is out of reach on a single card.

The VRAM Math

The 4080 Super edges the original 4080 on memory bandwidth (~736 vs ~717 GB/s), a few percent faster in practice; both are the same 16 GB fit class.

What Actually Fits (Model Picks)

Model	Quant	VRAM used	Speed	Notes
Qwen 3.5 9B	Q8_0	~10 GB	~40 tok/s	Highest-quality small model
gpt-oss 20B	Q4_K_M	~12-13 GB	~30 tok/s	Best OpenClaw agent pick
Qwen 3.6 27B	Q3_K_S	~14 GB tight	~25 tok/s	Fits but quality drops at Q3
Llama 3.1 8B	Q8_0	~9 GB	~45 tok/s	Fast general assistant

What You Can’t Run

Qwen 3.6 27B at Q4_K_M or higher — needs ~17-18 GB, over the 16 GB ceiling.
Any 70B model at a usable quant — a single 16 GB card is far short; you need 24 GB+ or two cards.
Long 128K context on a 20B model — the KV cache pushes you over 16 GB; cap context at 16-32K.

🎮 16 GB TODAY, OR STEP UP TO 24-32 GB

4080/4080 Super listings come and go; for a linkable 16 GB card the 4070 Ti Super is the same fit class. Want to run 27B at a good quant? Step up to a 24 GB 4090 or the 32 GB 5090.

4070TiSGIGABYTE RTX 4070 Ti Super 16 GB ↗ 4090GIGABYTE RTX 4090 24 GB ↗ 5090GIGABYTE RTX 5090 32 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 gpt-oss:20b
openclaw config set agents.defaults.models.chat "ollama/gpt-oss:20b"

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.