NVIDIA · Ada Lovelace

GeForce RTX 4070 SUPER

The reference specs, practical tradeoffs, and local models worth trying. A starting guide—not a benchmark or a promise that every workload fits.

GPU specs reviewed 2026-10-05 · Model listings reviewed 2026-10-05 · No live price or stock feed

VRAM: 12 GBReference GPU [9]
Memory: GDDR6XReference GPU [9]
Memory bus: 192-bitReference GPU [9]
Board power: 220 WReference board, not PSU [9]
Architecture: Ada LovelaceReference GPU [9]

Gaming & local AI

Consider for 1440p gaming; validate demanding ray-tracing and texture settings.

12 GB CUDA candidate. Choose the exact model, quantization and context first; equal VRAM does not mean equal inference speed. [1] [2]

Software first: Check your GPU architecture, NVIDIA driver and CUDA-capable runtime version. Blackwell cards need a runtime that supports them. [1] [2]

Older-generation option; verify the exact listing, seller, condition, warranty, cooling and connectors. No current price or availability checked. Board-partner specifications may differ.

Local models to try

These are capacity-screened candidates from our researched model library, not tested recommendations for this exact card. Published download size is not total VRAM required. Hardware/runtime support must be checked separately.

Start with an explicitly selected 4,096-token context, one request at a time. This is our suggested trial setup, not a guaranteed runnable context. Longer contexts and concurrent requests increase memory requirements. Model weights, KV cache, runtime allocations, and other GPU workloads all compete for memory. [24] [25]

Model / exact Ollama tag	Published download	Capacity screen	What to explore
Qwen3 4B`qwen3:4b` [27]4.02B parameters · Q4_K_M at review	2.5 GBDownload, not VRAM required	First trialsCapacity screen only; not tested.	Editorial shortlist candidate for multilingual chat, writing and instruction-following experiments; evaluate outputs for your task.Text input and text output; the tag page does not advertise vision or audio.
Gemma 3 4B`gemma3:4b` [32]4.3B parameters · Q4_K_M at review	3.3 GBDownload, not VRAM required	First trialsCapacity screen only; not tested.	Editorial shortlist candidate for text-and-image question answering and summarization experiments; verify visual interpretations.Text and image input, text output. Published 3.3 GB is the download listing, not a runtime VRAM requirement or usable-context guarantee.
Qwen2.5 Coder 7B`qwen2.5-coder:7b` [34]7.62B parameters · Q4_K_M at review	4.7 GBDownload, not VRAM required	First trialsCapacity screen only; not tested.	Editorial shortlist candidate for code generation, explanation and repair experiments; review and test generated code.Text/code input and text/code output; the tag page does not advertise vision or audio.
Qwen3 8B`qwen3:8b` [28]8.19B parameters · Q4_K_M at review	5.2 GBDownload, not VRAM required	First trialsCapacity screen only; not tested.	Editorial shortlist candidate for multilingual chat, writing and instruction-following experiments; evaluate outputs for your task.Text input and text output; the tag page does not advertise vision or audio.
DeepSeek R1 0528 Qwen3 8B`deepseek-r1:8b` [33]8.19B parameters · Q4_K_M at review	5.2 GBDownload, not VRAM required	First trialsCapacity screen only; not tested.	Editorial shortlist candidate for reasoning-oriented mathematics, programming and logic prompts; independently check conclusions.Text input and text output; the tag page does not advertise vision or audio.
Qwen3 14B`qwen3:14b` [29]14.8B parameters · Q4_K_M at review	9.3 GBDownload, not VRAM required	Tighter memoryMore context/overhead risk; may offload.	Editorial shortlist candidate for multilingual chat, writing and instruction-following experiments; evaluate outputs for your task.Text input and text output; the tag page does not advertise vision or audio.

How these candidates are selected—and what we do not know

Editorial heuristic, not a memory calculator: downloads up to 65% of nominal VRAM are labeled “First trials”; above 65% and up to 85% are “Tighter memory.” Larger downloads are omitted from this single-GPU shortlist. These percentages are deliberately simple screening rules, not measured allocation budgets; the source sizes are rounded downloads and do not establish exact GB/GiB memory fit.

We have not calculated model-specific KV cache sizes, measured peak VRAM, tested this card/model/backend combination, or established a safe maximum context. Parameter count and weights quantization alone cannot answer those questions. Equal-VRAM GPUs receive the same capacity screen, not the same speed or software-compatibility claim.

A model's advertised maximum context is a capability limit, not a promise that this card can allocate it. A model can launch with CPU offloading and still be unsuitable for your desired speed. Inspect the actual allocation before deciding it “runs well.” [24] [25]

Try it, then check the allocation

After installing a supported Ollama runtime and GPU driver, use this first-trial command. This example is a cautious starting experiment—not an assertion that the configuration has been tested here.

ollama run qwen3:4b

Inside the interactive Ollama session, explicitly set the trial context before sending a prompt. Do not rely on a default context setting. [25]

/set parameter num_ctx 4096

In another terminal, check the loaded model:

ollama ps

Inspect PROCESSOR and CONTEXT. CPU/GPU splitting means the trial is not fully GPU-resident. A “100% GPU” result is evidence for that running configuration, not for every future context or workload. [24]
Test the workload you actually need. If allocation fails or unwanted offloading appears, try a smaller model or reduce context/concurrency. Avoid simultaneous games, other models, or GPU-heavy applications during the trial.
Model weight quantization and KV-cache quantization are separate. Ollama documents f16 as its default cache type; cache quantization and supported Flash Attention can reduce memory usage, with model/backend support and potential quality tradeoffs. Do not assume they halve total model memory. [25]
Vision inputs can introduce additional allocations; a small text-only trial does not validate multimodal or long-context operation.

Sources & review notes

Manufacturer specs and exact model listing details are sourced below. Trial labels and workflow choices are GPUcrate editorial guidance. Model tags, downloads and runtime behavior can change; verify the listing and installed version before copying commands.