Home/Hardware/RTX 5070
RTX 5070 · 12GB GDDR7 on a 192-bit bus, 672 GB/s

Best AI Models for the RTX 5070

The 14B class on the fastest 12GB card on this hub: how much context 12GB really leaves, what it cannot hold, and the 8GB laptop variant. Footprints from Ollama’s tag sizes, speeds as bandwidth ceilings.

📅 Published: November 20, 2026🔄 Last Updated: November 2026✓ Manually Reviewed
Short answer

The best AI model for the RTX 5070 is Phi-4 14B — 9.1GB at Q4_K_M as Ollama lists it, with a bandwidth ceiling of ~74 tok/s (672 GB/s ÷ 9.1GB — a ceiling, not a measurement), the highest of any 12GB card on this hub. The 5070 runs the same 14B class as an RTX 3060 12GB or 4070, with more bandwidth: Qwen 3 14B (9.3GB) as the all-rounder, Qwen 2.5 Coder 14B and DeepSeek-R1 14B (9.0GB each) for code and reasoning, Gemma 4 12B (8.0GB) when you want more room for context, and Flux.1 dev FP8 for images. What 12GB cannot hold is the 16GB list: gpt-oss:20b (14GB) and Devstral Small 2 (15GB) are both out, and a 14B leaves under 3GB for context. We do not own a 5070; nothing here is a measured rate.

Models that run well on the RTX 5070

ModelSizeBuildMemory usedCeiling tok/s*
WINNERPhi-4 14B
The pick for math, logic and focused prompts, and the winner on this site’s 12GB tier page and its other 12GB card pages. Ollama lists a 16K context window for it; for long documents, Qwen 3 14B below has 40K.
ollama pull phi4
14B (dense)Q4_K_M~9.1GBceiling ~74 tok/s (672 ÷ 9.1GB)
Qwen 3 14B
The all-rounder, with a thinking mode and a 40K window. On 12GB the window is not the limit, the memory is: ~2.7GB spare is roughly 16K tokens of FP16 KV cache (see the notes).
ollama pull qwen3:14b
14B (dense)Q4_K_M~9.3GBceiling ~72 tok/s (672 ÷ 9.3GB)
Qwen 2.5 Coder 14B
The coding pick for 12GB: the largest dense coder that fits with room for a working context. The 24B coders need a 16GB card.
ollama pull qwen2.5-coder:14b
14B (dense)Q4_K_M~9.0GBceiling ~75 tok/s (672 ÷ 9.0GB)
DeepSeek-R1 14B
Visible chain-of-thought for hard problems. Reasoning traces eat context fast, and on 12GB that is the constraint to watch.
ollama pull deepseek-r1:14b
14B (dense)Q4_K_M~9.0GBceiling ~75 tok/s (672 ÷ 9.0GB)
Gemma 4 12B
Text and image input, a 256K window, and the smallest footprint of the 12–14B picks, which buys about a gigabyte more context room than the 14Bs. Its 13GB Q8 build is a 16GB-card option.
ollama pull gemma4:12b-it-q4_K_M
12B (dense)Q4_K_M~8.0GBceiling ~84 tok/s (672 ÷ 8.0GB)
Flux.1 dev FP8
Fits, just. Blackwell’s FP8 tensor cores make the FP8 build the fast path, but ~11GB of transformer leaves no room for the T5 text encoder to stay resident, so expect it to be swapped in and out between prompts. 16GB cards keep both loaded.
ComfyUI: flux1-dev-fp8.safetensors
12BFP8~11GBdiffusion — compute-bound, not timed
Llama 3.1 8B
The long-context and speed lane: ~7GB left over is enough for a genuinely long window, or for a second small model alongside.
ollama pull llama3.1:8b
8BQ4_K_M~4.9GBceiling ~137 tok/s (672 ÷ 4.9GB)

* Ceiling tok/s = the device’s vendor-listed memory bandwidth ÷ the model’s file size as Ollama lists it — the rate a dense model cannot exceed, because every generated token reads the whole weight set once. We have not measured this hardware. Real-world generation usually lands at 50–80% of the ceiling and falls as the context grows; mixture-of-experts rows read only their active parameters, so no ceiling is quoted for them, and image models are compute-bound, so no render time is.

What won't fit on the RTX 5070

  • ✗gpt-oss:20b (needs 14GB) — OpenAI’s open-weight model is sized for 16GB. On 12GB part of it lands in system RAM, and those layers run at system-RAM pace.
  • ✗Devstral Small 2 24B (needs 15GB at Q4_K_M) — The strongest coder in the 16GB class; on 12GB, Qwen 2.5 Coder 14B is the ceiling.
  • ✗Gemma 4 12B at Q8_0 (needs 13GB) — The near-lossless 12B is the first thing a 16GB card adds over this one.
  • ✗Qwen3.6 27B (needs 17GB at Q4_K_M) — A 24GB-card model; no 12GB card comes close.

RTX 5070-specific notes

  • →Where this page’s numbers come from: we do not own a 5070. Footprints are the file sizes Ollama prints on each tag page (read 20 Nov 2026), and every tok/s is a ceiling — NVIDIA’s listed 672 GB/s divided by the gigabytes a dense model reads per token. Real engines land below it and slow as the context fills. We quote no measured rate.
  • →What 12GB means for context, in numbers: Qwen 3 14B has 40 layers, 8 KV heads and a head dimension of 128, so its FP16 KV cache costs 2 × 40 × 8 × 128 × 2 bytes = 160KB per token. The ~2.7GB left after its 9.3GB of weights is about 16K tokens, before the CUDA runtime and your desktop take their share. Ollama’s FAQ documents two levers: Flash Attention (OLLAMA_FLASH_ATTENTION=1) and, with it on, a quantized KV cache (OLLAMA_KV_CACHE_TYPE=q8_0 uses about half the memory of the f16 default). On a 16GB card the same model has ~6.7GB spare, around 40K tokens at FP16.
  • →Bandwidth is where the 5070 earns its place among 12GB cards: 672 GB/s against 504 for the RTX 4070 and 360 for the RTX 3060 12GB, so its ceilings on the same 14B models are a third higher than a 4070’s and 87% higher than a 3060’s. The model list is identical on all three; only the ceiling moves.
  • →The laptop trap: NVIDIA’s laptop comparison lists the RTX 5070 Laptop GPU with either 12 GB or 8 GB of GDDR7 on a 128-bit bus, 4,608 CUDA cores and a 50–100 W range, against the desktop card’s 12 GB on 192-bit, 6,144 cores and 250 W. An 8GB 5070 laptop is an 8GB-tier machine where no 14B fits, and even the 12GB version has a narrower bus. Check the memory size in the laptop’s spec sheet, not the GPU name.
  • →RTX 5070 or RTX 5060 Ti 16GB is the real decision at this budget. The 5070 has 50% more bandwidth (672 vs 448 GB/s), so it runs the shared 14B list with a 50% higher ceiling; the 5060 Ti 16GB has 4GB more memory, which adds gpt-oss:20b, Devstral Small 2, Gemma 4 at Q8 and real context room. If your models are 14B and smaller, the 5070 is the faster card; if you keep running into the 12GB wall, more memory beats more speed.
  • →Power, from NVIDIA’s spec table: 250 W total graphics power, a 650 W required system power supply, and either 2× PCIe 8-pin cables through the adapter in the box or a single 300 W or greater PCIe Gen 5 cable. Prefer a native cable to an adapter chain and seat it fully.
  • →The Blackwell software trap: like every RTX 50-series card, the 5070 reports compute capability 12.0 (sm_120), and PyTorch wheels built for CUDA 12.6 or older have no kernels for it, so old pinned projects fail with “no kernel image is available for execution on this device”. Ollama ships its own CUDA runtime and lists the 5070 in its compute-capability 12.0 row; ComfyUI and projects with compiled extensions are where it bites.
  • →Price on 20 Nov 2026: NVIDIA announced the 5070 at $549. videocardprices.com’s last in-stock new reading was $1,010.92 at Amazon on 9 Oct 2026 (84% over MSRP; it had no newer in-stock price on 20 Nov), and resaleprices.com put the used average at $824 (fair asking range $782–836, 107 eBay listings). The same day the 16GB RTX 5060 Ti listings asked $850–880 and a used RX 9070 XT averaged $889: within 10% of this card’s used average, for 4GB more memory.

Quick start

Install Ollama
curl -fsSL https://ollama.com/install.sh | sh
Run the winner
ollama run phi4
Confirm it is fully on the GPU (must read 100% GPU)
ollama ps
Once your hardware is sorted

Go from "it runs" to actually building

All 561 chapters — running local models, RAG, agents, fine-tuning — plus the Python Lab and every course added later.

$149 once unlocks everything, forever — about $0.27/chapter for life. Prefer to spread it out? Pro is $79/year (saves 27%) or $8.99/month.
Secure checkout by Lemon Squeezy — your card never touches this siteInstant access the moment you payFirst chapter of every course is free — try before you buy

Frequently asked questions

What is the best LLM for an RTX 5070?

Phi-4 14B at Q4_K_M — 9.1GB as Ollama lists it, with a bandwidth ceiling of ~74 tok/s (672 ÷ 9.1GB). For a general all-rounder with a longer window, Qwen 3 14B (9.3GB); for code, Qwen 2.5 Coder 14B; for reasoning, DeepSeek-R1 14B (9.0GB each); for the most context room, Gemma 4 12B (8.0GB). They leave between ~2.7GB and ~4GB of the card for context. Every figure is a tag size or a ceiling, not a rate we measured.

Can the RTX 5070 run gpt-oss:20b?

Not fully on the GPU. Ollama’s gpt-oss:20b tag is 14GB, two more than the card has, so part of the model goes to system RAM and those layers run at system-RAM pace. It is a 16GB-card model: the RTX 5060 Ti 16GB, RX 9070 XT and RTX 5070 Ti all hold it. On the 5070 the 14B class is the honest ceiling, and a good one.

RTX 5070 or RTX 5060 Ti 16GB for local AI?

Speed versus capacity. The 5070’s 672 GB/s gives a 50% higher ceiling than the 5060 Ti’s 448 on every model both can hold, which is the whole 14B class. The 5060 Ti 16GB holds four things the 5070 cannot: gpt-oss:20b (14GB), Devstral Small 2 (15GB), Gemma 4 12B at Q8 (13GB), and a 14B with ~7GB of context instead of ~3GB. Buy the 5070 if your models are 14B and smaller; buy 16GB if you keep hitting the memory wall.

RTX 5070 vs RTX 4090 for AI: which is better?

The 4090, by a tier rather than a margin. Its 24GB holds Qwen3.6 27B (17GB) and the 32B coders (20GB) that no 12GB card can load, and its 1,008 GB/s is half again the 5070’s 672. What the 5070 has is Blackwell’s FP8 and FP4 tensor cores and 250 W of graphics power against 450 W. If your models fit 12GB, the 5070 runs them well; if you want the 27–32B class, the comparison is not close.

Do the RTX 5070’s 988 AI TOPS matter for local LLMs?

Less than the spec sheet suggests. NVIDIA lists 988 AI TOPS for the 5070, a tensor-core throughput figure. Generating tokens for one user is limited by memory bandwidth, because every token reads the whole model, so the number that predicts chat speed is the 672 GB/s, not the TOPS. Compute shows up in prompt processing (reading a long prompt or document before the first token) and in image generation.

Related guides

Other GPUs

Ready to Go Beyond Tutorials?

25 structured courses with hands-on chapters - build RAG chatbots, AI agents, and ML pipelines on your own hardware.

Was this helpful?

Free Tools & Calculators