Best Local LLM for RTX 5070 Ti (2026): gpt-oss 20B Wins

Bottom Line

• RTX 5070 Ti (16GB): run gpt-oss 20B. The official MXFP4 file is 11.28 GiB. llama.cpp’s gpt-oss guide measures 189 tok/s generation on a 5070 Ti.
• Want a stronger coder? Qwen3.8 27B fits at 3.5 bits (10.96 GiB, IQ3_S). It is dense, so it is much slower than gpt-oss 20B.
• RTX 5070 (12GB): run Qwen3.5 9B at Q8_0 (8.87 GiB) or Gemma 4 12B QAT (6.50 GiB). gpt-oss 20B only runs with some layers on the CPU.
• Laptop warning: the laptop RTX 5070 Ti has 12GB, not 16GB. NVIDIA’s laptop spec page says so. Use the 12GB section.
• Price check: a new 5070 Ti costs about $1,150-1,260 as of September 2026. That is close to a used 24GB RTX 3090. Read the value section before you buy one for LLMs.

Ready to buy? See the tested hardware list with current prices.

Best Local LLM for RTX 5070 Ti (16GB)

Sizes are exact Hugging Face file sizes, read 2026-09-29. Context memory comes on top of each file.

Why gpt-oss 20B wins on this card. It is a mixture-of-experts model with small active weights, so it generates fast on 896 GB/s. It also calls tools cleanly, which matters for agents. Its context is cheap: from its config.json, only 12 of its 24 layers use full attention, with 8 KV heads of 64 dimensions. That is about 24 KiB per token at f16, so 32K of context costs about 0.75 GiB (our arithmetic).

Can it hold the full 128K context? The KV cache at 128K is about 3.0 GiB. Weights plus cache come to about 14.3 GiB before llama.cpp’s compute buffers. That is borderline on 16GB. The llama.cpp guide caps 16GB cards at 32K in its example. We found no 5070 Ti test at 128K, so start at 32K and step up while you watch memory.

Why Qwen3.8 27B is the second pick. It is a dense 27B, so every token reads all 11 GiB of weights. At 896 GB/s the ceiling is about 76 tok/s; real llama.cpp speed is lower. That is our bandwidth estimate, not a measurement. In return you get the best coding model that fits in 16GB. Our Qwen by VRAM guide has the quality numbers for this file.

The MoE option. Qwen3.6 35B-A3B at UD-IQ3_S (12.74 GiB) fits fully. One community write-up (magnus919.com, May 2026) reports 121 tok/s on a 5070 Ti with a 10.88 GB community quant plus MTP speculative decoding. That is a different file and setup, so treat it as community-reported. At 4-bit (20.61 GiB) the model does not fit; you must move experts to system RAM with --n-cpu-moe. We found no measured 5070 Ti speed for that setup.

Is 16GB of VRAM Enough on the 5070 Ti?

Yes, for the models above. It is not enough for these:

• A 27B at 4-bit plus useful context. The 4-bit Qwen 27B files are 15.7-16.4 GiB before any context.
• Qwen3.6 35B-A3B at 4-bit without CPU offload (20.61 GiB).
• 70B models at any quality you would want to use.

If you need those, the next step is 24GB (a used RTX 3090) or 32GB (an RTX 5090).

Best Local LLM for RTX 5070 (12GB)

The 12GB card is a 9B-12B card. Qwen3.5 9B at Q8_0 is the quality pick; Gemma 4 12B QAT leaves the most room for context. There is no Qwen3.6 or Qwen3.8 model below 27B, so the newest small Qwen is still Qwen3.5.

gpt-oss 20B is 11.28 GiB, which leaves almost nothing for context on 12GB. One community run in llama.cpp’s gpt-oss thread put 2 MoE layers on the CPU (--n-cpu-moe 2) and measured 62 tok/s at 32K context on a 5070.

Speed: Bandwidth Decides

Community-reported llama.cpp CUDA scoreboard results (discussion #15013), Llama 7B Q4_0, flash attention on, tokens per second:

Generation speed follows memory bandwidth almost exactly. The 5070 Ti has 2.0x the 5060 Ti’s bandwidth and generates 1.95x faster. On gpt-oss 20B, llama.cpp’s own guide measures 189 tok/s on the 5070 Ti and 112 on the 5060 Ti (discussion #15396).

Specs: 16GB GDDR7 on a 256-bit bus and 300W for the 5070 Ti; 12GB on 192-bit and 250W for the 5070 (nvidia.com). Bandwidth is 896 and 672 GB/s (TechSpot launch coverage).

Is the 5070 Ti the Right 16GB Card?

Prices as of September 2026. They move weekly in the 2026 memory shortage.

Launch MSRPs were $749 (5070 Ti) and $549 (5070). Street prices now run roughly 55-110% above them.

The honest comparison nobody makes: at $1,150 or more, the 5070 Ti costs about what a used RTX 3090 costs. The 3090 has 24GB at 936 GB/s. That extra 8GB runs a 27B at 4-bit with context, which the 5070 Ti cannot. The 5070 Ti wins on warranty and a new card’s condition.

• You already own a 5070 Ti: keep it and run gpt-oss 20B at about 189 tok/s.
• You are buying for LLMs only: a used RTX 3090 24GB gives more memory for the same money. See how to buy one safely.
• You want 16GB for less: the RTX 5060 Ti 16GB loads the same models at about half the speed, for about $400 less.
• You want 32GB: the RTX 5090 runs a 27B at 8-bit, at a much higher price.

We do not carry a 5070 Ti link yet, so none is on this page.

OpenClaw Setup

Point OpenClaw at your local model through Ollama:

# pull gpt-oss 20B, 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 and cap context to what your memory holds. See the tool-calling reliability guide.

Before you order parts, check the tested hardware list for current prices by tier.

See Also

• Best Local LLM for 16GB VRAM — every 16GB card, and context math per length
• Best Qwen Model for 16GB VRAM — the Qwen3.8 27B IQ3_S file in detail
• Best Local LLM for RTX 5060 Ti 16GB — the cheaper 16GB Blackwell card
• Best Local LLM for RTX 4070 Ti Super (16GB) — the last-generation 16GB card
• Best Local LLM for RTX 4070 (12GB) — same 12GB fit class as the 5070
• Is a Used RTX 3090 Worth It? — the 24GB card at the same price
• Best Local LLM for RTX 5090 (32GB) — the Blackwell step up
• Best Local LLM by GPU (hub)