Best local LLM for rag & documents on a NVIDIA GeForce RTX 4090 (2026)
All figures assume an f16 KV cache, a 0.6 GB display reserve on the GPU, and 64 GB of DDR5 system RAM for the offload tiers. Tune these in the calculator.
The verdict
Qwen3.5 35B A3B Q4_K_M
Fits on GPU at 16K · RAG & documents score 9/10 · ≈ 184 tok/s
The meta pick, full stop. Near-dense-30B quality at 3B-active speed, and expert offload puts it on 8 GB cards.
llama-server -m Qwen3.5-35B-A3B-Q4_K_M.gguf -c 16384 --flash-attn -ngl 99
Worthy alternates
Qwen3.5 9B Q8_0
Fits on GPU · ≈ 69 tok/s · RAG & documents 8/10
The new small default. Frontier-distilled, natively multimodal, and embarrassingly good for 6 GB of weights.
Gemma 4 26B A4B Q5_K_M
Fits on GPU · ≈ 139 tok/s · RAG & documents 8/10
Google's fast MoE with native audio in. Nearly all of its weight sits in routed experts, so expert offload runs it comfortably on 12 GB cards.
Tune this for your exact RAM and settings in the calculator → · All models on the NVIDIA GeForce RTX 4090
Frequently asked questions
What is the best local LLM for rag & documents on a NVIDIA GeForce RTX 4090?
Qwen3.5 35B A3B at Q4_K_M — it scores 9/10 for rag & documents and runs as "Fits on GPU" at 16K context on the NVIDIA GeForce RTX 4090.
How much context do I need for rag & documents?
We recommend 24K tokens for rag & documents (minimum 12K). These picks are computed at 16K.
How fast will it run on a NVIDIA GeForce RTX 4090?
Roughly 184 tokens/sec for Qwen3.5 35B A3B — fast for interactive use.
Do I need more than 24 GB of VRAM for rag & documents?
No — the pick above needs 23.6 GB of VRAM at 16K.
What settings should I use?
Start with our command: llama-server -m Qwen3.5-35B-A3B-Q4_K_M.gguf -c 16384 --flash-attn -ngl 99 — then tune context and KV quant in the fit calculator.