Is a used RTX 3090 still worth it for local LLMs in November 2026?

Yes, if you run GGUF models in llama.cpp and can live without FP8 and FP4. On 7 November 2026 a used RTX 3090 asked a median 1,790 euros on Kleinanzeigen, against about 3,000 euros for a used 4090 and 6,000 euros for a 5090, and it still has 24 GB at 936 GB/s. It generates 161.8 tok/s on gpt-oss-20b in llama.cpp (4090: 222.0, 5090: 282.5), which makes it the cheapest of the three per token per second. What it lacks is newer compute: no FP8 or FP4 tensor cores, and prompt processing at roughly half a 5090's rate.

The guide's build 6210 dates from August 2025; newer llama.cpp builds (b12189 on 7 November 2026) may be faster on every card. Dates in the right column are when the page was read or, for benchmarks, when the run was published.

What does a used RTX 3090 cost in November 2026?

About 1,550 to 1,850 euros in Germany for a working card, well above the 12-month low of 1,200 euros that gpuprix recorded for German retail offers. I read the first results page of Kleinanzeigen for "rtx 3090" on 7 November 2026: six working cards were listed between 1,550 and 2,100 euros, all but one marked negotiable, and one card with "a broken RAM chip" was offered at 1,000 euros. eBay's sold-listings pages refused automated access when I checked, so the table below uses asking prices, which sit above what cards actually sell for.

Sources: Kleinanzeigen RTX 4090 and RTX 5090 search pages, read 7 November 2026, complete PCs, laptops, wanted ads and swaps excluded. gpuprix pages for the 3090 in Germany and 4090 in the US, last updated 4 November 2026; they track retailer offers, new and renewed, not private sales. Prices include whatever VAT the seller charges; private sales carry none.

Retail is no cheaper. gpuprix's cheapest renewed 3090 in the US was $1,795 on Amazon, and its 3-month median across tracked offers was $1,999; in Germany it was 1,990 euros. The same tracker had the cheapest in-stock new RTX 5090 at $7,550 in the US on 4 November (gpuprix), and the 2026 memory shortage has dragged older 24 GB cards up with it.

How fast is an RTX 3090 for LLMs?

Fast enough for one person on any model that fits: 110 to 162 tok/s on 20-35B mixture-of-experts models and about 35 tok/s on a dense 27-32B model at 4-bit, all from llama.cpp. Decode speed follows memory bandwidth, and 936 GB/s is about 52% of a 5090's 1,792 GB/s (gpuprix spec tables).

tok/s, single stream. Hardware Corner names llama.cpp build 8189 on its RTX 5090 page and no build on the 3090 page, so the 3090 and 5090 rows may come from different builds. The 4090 Qwen3 32B figure is from Hardware Corner's 4090 page.

The dense rows show the bandwidth ceiling at work. Qwen3 32B at Q4 is 18.6 GB per Naveen Kulandaivelu's model table, and a dense model reads all of it per token, so the ceiling is 936 / 18.6 = 50 tok/s on a 3090 and 1,792 / 18.6 = 96 tok/s on a 5090 (my arithmetic). The reported 35.1 and 61.4 tok/s are 70% and 64% of those ceilings. The 4090, at 1,008 GB/s per gpuprix, is only 13% faster than the 3090 on this model.

Prompt processing is where the age shows. Prefill is compute-bound, and on gpt-oss-20b at 32k tokens the guide has the 3090 at 3,577 tok/s against 6,291 for the 5090. By my arithmetic, a 30,000-token document takes about 8.4 seconds to read on the 3090 and 4.8 seconds on the 5090. For chat that hardly matters; for RAG over long files or coding agents that resend big contexts, it does.

What does missing FP8 and FP4 mean on Ampere?

It means FP8 and NVFP4 checkpoints run, but as weight-only formats that save memory without the speedup Blackwell owners get. vLLM's FP8 page says FP8 computation needs compute capability 8.9 or higher (Ada, Hopper, Blackwell); on Ampere, which vLLM's quantization table lists as SM 8.0/8.6, FP8 models run as W8A16 through Marlin kernels. For NVFP4, vLLM's ModelOpt page says GPUs without a native FP4 GEMM fall back to W4A16 via Marlin and log a warning that "this may reduce throughput for compute-heavy workloads".

MXFP4 is the friendly case (GPUwerk's quantization notes explain the format). The vLLM gpt-oss recipe says gpt-oss works on Ampere by default with the TRITON_ATTN backend and Marlin MXFP4 MoE kernels, and llama.cpp runs the MXFP4 GGUF on a 3090 at the 161.8 tok/s above. Decoding at batch 1 is bandwidth-bound anyway, so weight-only formats cost little there. The loss shows up in prefill and in batched serving, where Ada and Blackwell multiply in FP8 (and Blackwell in FP4) while the 3090 dequantizes to FP16 first.

The bigger risk is software direction. New releases increasingly ship Blackwell-first. NVIDIA's Nemotron 3.5 Bolt 74B-A9B 24GB (15 October 2026) is packed to fit 24 GB, which sounds made for a 3090, but it stores its Mamba layers in NVFP4, documents vLLM serving on a nightly build with an extra kernel package, and lists A100, H100 and Blackwell as targets without mentioning the 3090. Check the model card before you plan around a new model. The planned piece on NVFP4, MXFP4 and FP8 covers the formats themselves.

What fits in 24 GB?

Every current 20-35B mixture-of-experts model at 4-bit, and dense 27B models with moderate context. The table adds the KV cache to the weight file, using 2 x attention layers x KV heads x head dimension x 2 bytes per token from each model's config.json (FP16 cache), against about 23 GiB of usable memory.

File sizes from the Hugging Face file listings of ggml-org/gpt-oss-20b-GGUF, unsloth/gemma-4-26B-A4B-it-GGUF, unsloth/Qwen3.6-35B-A3B-GGUF, unsloth/Qwen3.8-27B-GGUF and ggml-org/gpt-oss-120b-GGUF, read 7 November 2026. Layer counts from the gpt-oss-20b, Qwen3.8-27B, Qwen3.6-35B-A3B and Gemma 4 26B-A4B config files; only full-attention layers grow with context (the Qwen models' Gated DeltaNet layers keep a fixed state, Gemma's sliding layers cap at 1,024 tokens). Gemma's config sets attention_k_eq_v, so its real global cache may be half my figure. Totals exclude compute buffers of a few hundred MB.

Two cards (48 GB) take Qwen3.6-35B-A3B at Q8_0 (36.9 GB) or a dense 27B at Q8_0 with long context. gpt-oss-120b needs either three cards or llama.cpp's --n-cpu-moe, which keeps some expert weights in system RAM; the sibling article on MoE CPU offload has the RAM bandwidth math. For a quick fit check of your own model and context, the memory calculator uses the same formula.

Do two or four 3090s help, and is NVLink worth it?

More cards add capacity, not single-stream speed, and the NVLink bridge is optional. In XiongjieDai's llama.cpp benchmark repo (May 2024 snapshot), Llama 3 70B Q4_K_M ran at 16.29 tok/s on two 3090s and 16.89 tok/s on four, while Llama 3 8B Q4_K_M dropped from 111.74 tok/s on one card to 104.94 on four. Layer split runs the cards one after another, so each token still waits on every card in turn.

Tensor parallelism in vLLM does run the cards together. Paralon Cloud split Qwen3.8-27B (W4A16) across two RTX 4090s on plain PCIe, with no NVLink and no peer-to-peer, and reported 113 tok/s per request at concurrency 1 and 654 tok/s total at 32; the authors report that the all-reduce over PCIe never became the bottleneck at any concurrency they ran (Paralon, 9 September 2026). The 3090 supports an NVLink bridge, per NVIDIA's spec page; the 5090's spec page lists NVLink as "No". I found no current controlled test of what the bridge adds for inference, so I would not pay extra for one.

Power is the practical limit. Four cards are 1,400 W of board power before CPU and drives, and NVIDIA asks for a 750 W supply for a single 3090. The existing comparison of a DGX Spark against building a workstation covers what multi-GPU builds cost around the cards.

Is a used 3090 better value than a 4090 or 5090?

Per euro, yes, by a clear margin at November 2026 used prices. The calculation uses the llama.cpp guide's gpt-oss-20b generation speeds and the Kleinanzeigen median asking prices from the first table.

My arithmetic: 161.8 / 1.79 = 90.4; 222.0 / 3.0 = 74.0; 282.5 / 6.0 = 47.1; 24 / 1.79 = 13.4, 24 / 3.0 = 8.0, 32 / 6.0 = 5.3. Speeds from llama.cpp build 6210; prices are asking prices for used cards on Kleinanzeigen.

The 5090 is still the faster card by 75% on decode and nearly 2x on prefill, and it has 8 GB more. But two used 3090s (48 GB, about 3,600 euros) cost less than one used 5090 here. The 4090 sits awkwardly in the middle: the same 24 GB as a 3090, 37% faster on this model, for 68% more money.

Should you buy one?

For a single user running GGUF models up to about 35B MoE or 27B dense, I would buy a used 3090 in November 2026, and a second one later if 48 GB becomes the limit. Test it before paying: load a model that fills the memory and let it generate for ten minutes. The 1,000-euro listing with a broken memory chip on the same results page is a reminder of why. The decision steps in choosing a GPU for open-source LLMs (smallest model that does the job first, then card) apply here too.

Skip it if your work is batched serving with FP8 or NVFP4 checkpoints in vLLM, if you feed it 50k-token prompts all day, or if you want a warranty. Those cases favor Blackwell, and the comparison of the RTX 5090 with a DGX Spark covers the 32 GB and 128 GB options. If you only need to find out whether FP4 kernels or 32 GB change your results, GPUwerk's RTX 5090 rental runs at $0.89 an hour, which is cheaper than finding out with a purchase. If 24 GB is not enough and a 5090 is too expensive, the modded 48 GB RTX 4090 is the grey-market option; its risks are in RTX 4090 48 GB. A two- or four-card 3090 rig is set against a DGX Spark in DGX Spark vs an RTX 3090 rig.

FAQ

Sources

Every figure above is attributed to its source in the sentence or table where it appears. Pages were read on 7 November 2026 unless a date says otherwise.

• RTX 3090 listings, Kleinanzeigen, 2026-11-07
• RTX 4090 listings, Kleinanzeigen, 2026-11-07
• RTX 5090 listings, Kleinanzeigen, 2026-11-07
• GeForce RTX 3090 price in Germany, gpuprix, 2026-11-04
• GeForce RTX 3090 price in United States, gpuprix, 2026-11-04
• GeForce RTX 4090 price in United States, gpuprix, 2026-11-04
• GeForce RTX 5090 price in United States, gpuprix, 2026-11-04
• Quantization: supported hardware, vLLM documentation
• GeForce RTX 3090 Ti and 3090 specifications, NVIDIA
• GeForce RTX 5090 specifications, NVIDIA
• Performance of llama.cpp on Nvidia CUDA (discussion #15013), ggml-org/llama.cpp on GitHub, 2025-08-01
• guide: running gpt-oss with llama.cpp (discussion #15396), ggml-org/llama.cpp on GitHub, 2025-08-18
• RTX 3090 local LLM benchmarks, Hardware Corner, 2026-03
• RTX 5090 local LLM benchmarks, Hardware Corner, 2026-03
• FP8 W8A8 (LLM Compressor), vLLM documentation
• NVIDIA Model Optimizer (ModelOpt) quantization, vLLM documentation
• GPT-OSS recipe, vLLM recipes
• NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB model card, NVIDIA on Hugging Face, 2026-10-15
• GPU-Benchmarks-on-LLM-Inference, XiongjieDai on GitHub, 2024-05
• Two RTX 4090s over PCIe: splitting a 27B model beat two replicas by 2 to 4x, Paralon Cloud, 2026-09-09
• 96GB VRAM on a single GPU: benchmarking NVIDIA's RTX PRO 6000, Naveen Kulandaivelu, 2025-08-07
• gpt-oss-20b-GGUF files, ggml-org on Hugging Face
• gpt-oss-120b-GGUF files, ggml-org on Hugging Face
• gemma-4-26B-A4B-it-GGUF files, Unsloth on Hugging Face
• Qwen3.6-35B-A3B-GGUF files, Unsloth on Hugging Face
• Qwen3.8-27B-GGUF files, Unsloth on Hugging Face
• gpt-oss-20b config.json, OpenAI on Hugging Face
• Qwen3.8-27B config.json, Qwen on Hugging Face