Local LLM on NVIDIA RTX 4090

Is the NVIDIA RTX 4090 the right card for this?

24 GB is the capacity most local-LLM tooling is written around, and this card sits in that bracket at 1008 GB/s. Released in 2022, it is 4 years old in 2026 and sells for about $1600, or roughly $67 per gigabyte of VRAM. It draws up to 450 W, so the power supply and the case airflow are part of the budget, not an afterthought. Catalogue throughput class: Q4 7B: 55-75, Q4 32B: 18-25, Q4 70B: OOM (offload possible).

2 other cards here carry 24 GB, so all of them load the same weights and differ only in how fast they read them. At 1008 GB/s this is the quickest of that group, so what you gain over the others is tokens per second, not a model you could not otherwise open.

Released in 2022, this is one of the older entries in the guide, and age here is mostly an advantage: the current generation of local-AI tooling was written while this hardware was already widespread, so backend support is settled and the failure modes are documented. What it cannot have is any of the memory technology that arrived after it, and memory is what sets generation speed.

What the price gap buys

The cheapest 24 GB option in this guide is the NVIDIA RTX 3090 at about $700. This card costs $900 more and reads memory 8% faster, which is roughly the gain in tokens per second. It buys no additional model: both load exactly the same weights, so the question is only whether you are waiting on the output.

Released in 2022, this sits between the shelf and the second-hand market: new stock still turns up, and used examples are common enough that the price above spans both. That makes it worth checking the two markets against each other before committing, because the gap can be wider than the gap to the next card up.

Where the NVIDIA RTX 4090 sits in this guide

Across the 20 cards catalogued here, the NVIDIA RTX 4090 ranks 5th on memory bandwidth at 1008 GB/s, 12th cheapest at about $1600, and 17th on bandwidth per watt at 2.2 GB/s per watt. Those three positions, not the capacity figure, are what separate it from other 24 GB parts.

GDDR6X uses PAM4 signalling to push more bits per clock than plain GDDR6, which is where this card's 1008 GB/s comes from. It also runs hot: the memory modules, not the core, are usually the first thing to throttle under a long generation run.

This is a CUDA card, which in practice means every local runner in our tool directory supports it first. Ollama and llama.cpp pick it up without configuration; vLLM and TensorRT-LLM target it specifically. Driver and CUDA toolkit versions are the usual source of trouble, not the card.

Which catalogue models fit on NVIDIA RTX 4090?

Of the 46 models in our catalogue with per-size memory figures, 38 have at least one size that fits in 24 GB at Q4_K_M with room for context. The largest of each are below, biggest first.

Model	Largest size that fits	Weights at Q4_K_M
Qwen 2.5	32B	20 GB
Qwen 2.5-Coder	32B	20 GB
Qwen 3	32B	20 GB
DeepSeek R1	32B	20 GB
DeepSeek Coder	33B	20 GB
Gemma 4	31b	20 GB
Code Llama	34B	20 GB
QwQ	32B	20 GB

A further 30 smaller models also fit; see the model catalogue.

The next sizes up are out of reach without offloading: LLaVA 34B at 22 GB, Command R 35B at 22 GB, Llama 3.1 70B at 42 GB.

Expected tokens per second on NVIDIA RTX 4090

Memory bandwidth, not compute, is the bottleneck for single-stream token generation. The NVIDIA RTX 4090's 1008 GB/s puts the following ceiling on a batch of one at 2048 tokens of context.

Model	Quantization	Status	Approx tokens/sec
Llama 3.2 1B	Q4_K_M	Fits in VRAM	~181 tok/s
Llama 3.1 8B	Q4_K_M	Fits in VRAM	~60 tok/s
Llama 3.1 8B	Q8_0	Fits in VRAM	~45 tok/s
Mistral Nemo 12B	Q4_K_M	Fits in VRAM	~40 tok/s
Qwen 2.5 14B	Q4_K_M	Fits in VRAM	~35 tok/s
Qwen 3 32B	Q4_K_M	Fits in VRAM	~18 tok/s
DeepSeek R1 distilled 32B	Q4_K_M	Fits in VRAM	~18 tok/s
Llama 3.1 70B	Q4_K_M	Needs CPU offload	~3 tok/s

These come from a bandwidth heuristic (tokens/sec ≈ bandwidth in GB/s × a per-model efficiency factor), not from a test rig in our office. Real numbers move with model architecture, batch size and KV-cache size. Treat them as an order of magnitude, and treat the "reported throughput class" row in the table above as the community-measured figure.

Build recommendations for NVIDIA RTX 4090

Best model to download first

Llama 3.1 8B for everyday chat and coding; a 32B at Q4_K_M for top quality within 24 GB. Start with ollama pull llama3.1:8b.

Recommended inference backend

Use Ollama for general use or Mullama if you need a drop-in Ollama alternative with native bindings for 6 languages. For production serving on multi-GPU setups, see vLLM.

Sources

[1] nvidia.com/rtx-4090 · [2] community benchmarks · VRAM figures for catalogue models from src/data/models.json (2026-06-29); card specifications from src/data/gpus.json (2026-06-29).