What is the best local LLM for an RTX 5070 Ti?

gpt-oss-20b at Q4 if you want it all on the GPU: about 13 GB and roughly 249 tokens per second by our estimate. With 32GB of system RAM, Gemma 4 26B A4B at Q4 is the higher-quality pick.

How much VRAM does the RTX 5070 Ti have?

16 GB of GDDR7 with 896 GB/s of bandwidth, per NVIDIA. About 15 GB of that is usable for model weights and context by our estimate.

Can an RTX 5070 Ti run a 30B model?

Yes if it's a mixture-of-experts model like Qwen3.6 35B A3B, split between VRAM and system RAM. A dense model such as Qwen3.8 27B also loads that way but runs at only about 9 tokens per second by our estimate.

Is the RTX 5070 Ti good for local AI?

Yes for its price. It fits the same models as any 16GB card and runs them about three times faster than an RTX 4060 Ti 16GB (our estimate), but 16GB caps how big a dense model you can keep on the card.

Does 32GB of system RAM help an RTX 5070 Ti run LLMs?

Yes, for mixture-of-experts models. You keep the expert weights in RAM and the rest on the GPU, which lets 30B to 40B MoE models run at 50+ tokens per second by our estimate.

RTX 5070 Ti or RTX 5080 for local LLMs?

The 5070 Ti. Both have 16GB, so they run the same models, and the 5080's 960 GB/s is only about 7% more bandwidth than the 5070 Ti's 896 GB/s.

Best local LLM for RTX 5070 Ti: what fits in 16GB

Updated 2026-11-09. From NVIDIA's spec page, official model cards and our local finder's estimates.

Card	VRAM	Bandwidth	gpt-oss-20b tok/s	Gemma 4 12B tok/s
RTX 4060 Ti 16GB	16 GB	288 GB/s	~80	~14	Open in finder
RTX 4070 Ti Super	16 GB	672 GB/s	~187	~32	Open in finder
RTX 5070 Ti	16 GB	896 GB/s	~249	~42	Open in finder
RTX 5080	16 GB	960 GB/s	~267	~45	Open in finder

Which models fit entirely in the RTX 5070 Ti's 16GB?

Anything up to about 15 GB including the conversation. The card has 16 GB of GDDR7 per NVIDIA's spec page, and our finder keeps about 1 GB back for the display and drivers (our estimate). That rules out dense models much past 14B at good quality, but it leaves room for one strong mixture-of-experts model.

Model	Use	Quant	Download	tok/s
gpt-oss-20b OpenAI, 21B total, 3.6B active	Chat and writing	Q4	13 GB	~249 fast
Gemma 4 12B Google, 11.95B total	Chat and writing	Q8	13 GB	~42 fast
Ministral 3 14B Mistral, 14B total	Chat and writing	Q5	10 GB	~54 fast
Phi-4 14B Microsoft, 14.7B total	Chat and writing	Q5	10 GB	~51 fast
gpt-oss-20b OpenAI, 21B total, 3.6B active	Coding	Q4	13 GB	~249 fast
Gemma 4 E4B Google, 8B total, 4.5B active	Small and fast	Q8	8 GB	~112 fast
Ministral 3 8B Mistral, 8B total	Small and fast	Q8	8 GB	~63 fast
Qwen3 8B Qwen, 8.2B total	Small and fast	Q8	9 GB	~62 fast
Qwen3 4B Qwen, 4B total	Small and fast	Q8	4 GB	~126 fast

Our finder's top picks per use, highest-rated first. Q8 is about 8.5 bits per weight, Q5 5.7 and Q4 4.8. Fast is 30+ tokens per second, usable is 12+. All speeds are our estimates.

Bigger models with 32GB of system RAM

A lot of 5070 Ti builds pair the card with 32GB of DDR5, and that opens up a second tier. Mixture-of-experts models only use a small slice of their weights per token, so you can park most of the expert weights in system RAM and keep the rest on the GPU. In llama.cpp that's the --cpu-moe and --n-cpu-moe options; LM Studio and Ollama build on the same engine.

Model (Q4)	Size	Share in RAM	tok/s
Recursive Flint 40B A2.5B Recursive, 39.8B total, 2.5B active	24 GB	43%	~51 fast
Gemma 4 26B A4B Google, 25.2B total, 3.8B active	15 GB	11%	~95 fast
Qwen3 Coder 30B A3B Qwen, 30.5B total, 3.3B active	18 GB	26%	~58 fast
Qwen3.6 35B A3B Qwen, 35B total, 3B active	21 GB	36%	~50 fast
LFM2.5 24B A2B Liquid AI, 24B total, 2.45B active	14 GB	6%	~195 fast

Our estimate. Inputs: 15 GB usable VRAM at 896 GB/s, about 19 GB of the 32GB RAM free for weights at 60 GB/s (the system RAM speed our finder uses), and active weights read from each in proportion to where they sit. Faster DDR5 kits and smarter expert placement can beat these numbers; long contexts make them worse.

Dense models are a different story. Qwen3.8 27B at Q4 is about 17 GB, so roughly 19% of it lands in system RAM, and every token has to read all of it. That works out to about 9 tokens per second by our estimate. It runs, but you'll be waiting.

Which one should you pick?

Everyday chat and quick coding help: gpt-oss-20b at Q4. It stays on the card, so it's the fastest good option by a wide margin.
Best chat quality with 32GB RAM: Gemma 4 26B A4B at Q4, split across GPU and RAM, about 95 tokens per second (our estimate).
Coding agent work: Recursive Flint 40B A2.5B at Q4, about 51 tokens per second with the same split (our estimate). See our Recursive Flint guide for how it compares.
A dense model that fits: Gemma 4 12B at Q8, about 42 tokens per second.

If you're choosing between cards, the 5070 Ti runs the same models as the RTX 4060 Ti 16GB at about three times the speed, and it's within about 7% of the RTX 5080 on bandwidth. NVIDIA lists the 5070 Ti from $749. The next real step for local AI is more VRAM, not more speed: a 24GB or 32GB card such as the RTX 4090 or RTX 5090 fits 27B to 31B dense models without spilling.

How to run a local LLM on the RTX 5070 Ti

LM Studio is the easiest start on Windows: search the model, pick the quant from the tables above, and it uses the GPU automatically. Ollama does the same from the command line. For the RAM split on bigger MoE models, llama.cpp's server gives you the most control.

Keep the context window modest when a model is close to the 16GB line, because the conversation needs VRAM too. And if a model feels slow, try a smaller one before dropping below Q4.

Wondering whether local beats paying per token? Our cheapest LLM API guide and API price table show what the same work costs in the cloud.

Best local LLM for an RTX 5070 Ti

RTX 5070 Ti vs other 16GB cards for local AI

Which models fit entirely in the RTX 5070 Ti's 16GB?

Bigger models with 32GB of system RAM

Which one should you pick?

How to run a local LLM on the RTX 5070 Ti

Sources

FAQ

Related

More guides