Best value GPU for local LLMs in Europe (tok/s per euro)

Which card gives the most measured generation speed for each euro, at today's German retail prices? We divide real tokens/sec from sourced benchmarks by the lowest new price on billiger.de (incl. VAT, approx. price as of 2026-10-16).

Formula: value = measured tok/s ÷ GPU price (€) × 100 = tokens/sec you get per €100 of GPU. Higher is better. Where a source reports a range, we use the low end. GPU price only: CPU, RAM, PSU and case are not included. Used cards are often much cheaper.

Top value pick per model

Model	Best tok/s per euro	Fastest measured
Qwen3.8 27B	2× RTX 5070 Ti: 9.7 tok/s per €100 (230 tok/s, ≈€2,375)	2× RTX 5070 Ti: 230 tok/s (≈€2,375)
gpt-oss 20B	RTX 3060†: 19.6 tok/s per €100 (83.96 tok/s, ≈€429)	RTX 5090: 324.2 tok/s (≈€6,199)
Qwen3.6 35B-A3B	Radeon AI Pro R9700: 6.9 tok/s per €100 (117.69 tok/s, ≈€1,718)	RTX 5090: 258 tok/s (≈€6,199)
Qwen3.6 27B	Radeon AI Pro R9700: 3.8 tok/s per €100 (66.0 tok/s, ≈€1,718)	RTX 5090: 75.7 tok/s (≈€6,199)
Qwen3.8 Flash-Next	2× RTX 5070 Ti†: 7.7 tok/s per €100 (182 tok/s, ≈€2,375)	2× RTX 5070 Ti: 182 tok/s (≈€2,375)
Gemma 4 12B	RTX 3060: 9.0 tok/s per €100 (38.5 tok/s, ≈€429)	RTX 5090: 142 tok/s (≈€6,199)
Gemma 4 31B	RTX 5090: 1.1 tok/s per €100 (70.0 tok/s, ≈€6,199)	RTX 5090: 70.0 tok/s (≈€6,199)
Gemma 4 26B-A4B	RTX 4090: 4.1 tok/s per €100 (164 tok/s, ≈€3,955)	RTX 5090: 235 tok/s (≈€6,199)
Qwen3.5 9B	RTX 4090: 3.5 tok/s per €100 (137 tok/s, ≈€3,955)	RTX 5090: 208 tok/s (≈€6,199)
LFM2.5 24B-A2B	RTX 3060†: 29.1 tok/s per €100 (125 tok/s, ≈€429)	RTX 4090: 261-339 tok/s (≈€3,955)
Flint 40B-A2.5B	RTX 4090†: 4.5 tok/s per €100 (179.4 tok/s, ≈€3,955)	RTX 4090: 179.4 tok/s (≈€3,955)

Speeds differ in quant, engine and context length between sources (shown in each row), so treat small gaps as noise and look at the order of magnitude. Every row links its build page and original source. † = model is bigger than the card's VRAM at Q4_K_M, see the note under that model's table.

Qwen3.8 27B: tokens/sec per euro

GPU	VRAM GB	Measured tok/s	GPU price DE (new)	tok/s per €100	€ per tok/s
2× RTX 5070 Ti	32	230 capy-flash, 2 GPUs, MTP, decode at 8k prompt, 3-bit GSQ	≈€2,375	9.7	€10.3	build · source
2× RTX 5070 Ti	32	139 capy-flash (1 GPU), MTP, decode at 8k prompt, 3-bit GSQ	≈€2,375	5.9	€17.1	build · source
RTX 5060 Ti 16GB + RTX 3060	28	51.4 llama.cpp master (Oct 2026), UD-IQ4_XS on the 5060 Ti + MTP draft on the 3060 (n_max=2)	≈€1,168	4.4	€22.7	build · source
Radeon AI Pro R9700	32	57 llama.cpp Vulkan, Q4_K_M + MTP speculative decoding, reasoning on	≈€1,718	3.3	€30.1	build · source
RTX 5060 Ti 16GB + RTX 3060	28	37.6 llama.cpp master (Oct 2026), Q4_K_M, split + MTP draft (n_max=2)	≈€1,168	3.2	€31.1	build · source
RTX 5060 Ti 16GB + RTX 3060	28	28.0 llama.cpp master (Oct 2026), UD-IQ4_XS on the 5060 Ti alone, 8K ctx, q8 KV	≈€1,168	2.4	€41.7	build · source
RTX 5060 Ti 16GB + RTX 3060	28	21.2 llama.cpp (May 2026 build), Q4_K_M, layer split 5/3 across both GPUs	≈€1,168	1.8	€55.1	build · source
Radeon AI Pro R9700	32	30 Ollama, Q4_K_M	≈€1,718	1.7	€57.3	build · source
Radeon AI Pro R9700	32	22 vLLM (radiance), AWQ INT4, single request	≈€1,718	1.3	€78.1	build · source
RTX 5090	32	75.2 llama.cpp 0.7.0-dev (b11669), Q4_K_M (EPYC 9534 host)	≈€6,199	1.2	€82.4	build · source
RTX 3060	12	5.2-6.7 llama.cpp, llama-bench tg128 (quant per result)	≈€429	1.2 †	€82.5	build · source

† Qwen3.8 27B weights are about 16.5 GB at Q4_K_M / native size, more than this card's VRAM. The source may have used a smaller quant or partial CPU offload; check it before buying.

All measured speeds for Qwen3.8 27B: best rig for Qwen3.8 27B.

gpt-oss 20B: tokens/sec per euro

GPU	VRAM GB	Measured tok/s	GPU price DE (new)	tok/s per €100	€ per tok/s
RTX 3060	12	83.96 llama.cpp	≈€429	19.6 †	€5.1	build · source
RTX 5070 Ti	16	189.45 llama.cpp, MXFP4	≈€1,188	15.9	€6.3	build · source
RTX 5060 Ti 16GB	16	111.51 llama.cpp, MXFP4	≈€739	15.1	€6.6	build · source
RTX 5080	16	204.85 llama.cpp, MXFP4	≈€1,599	12.8	€7.8	build · source
Radeon AI Pro R9700	32	163.26 llama.cpp	≈€1,718	9.5	€10.5	build · source
RTX A4000	16	93.55 llama.cpp	≈€1,050	8.9	€11.2	build · source
RTX 4090	24	279.65 llama.cpp	≈€3,955	7.1	€14.1	build · source
RTX 5090	32	324.2 llama.cpp 0.7.0-dev, reported as Q4_K_M	≈€6,199	5.2	€19.1	build · source
RTX 4090	24	203 llama.cpp, llama-bench tg128 (quant per result)	≈€3,955	5.1	€19.5	build · source
RTX 5090	32	313 llama.cpp, llama-bench tg128 (quant per result)	≈€6,199	5.0	€19.8	build · source
RTX 5090	32	282.51 llama.cpp, MXFP4	≈€6,199	4.6	€21.9	build · source

† gpt-oss 20B weights are about 13.0 GB at Q4_K_M / native size, more than this card's VRAM. The source may have used a smaller quant or partial CPU offload; check it before buying.

All measured speeds for gpt-oss 20B: best rig for gpt-oss 20B.

Qwen3.6 35B-A3B: tokens/sec per euro

GPU	VRAM GB	Measured tok/s	GPU price DE (new)	tok/s per €100	€ per tok/s
Radeon AI Pro R9700	32	117.69 llama.cpp Vulkan, Q4_K_M	≈€1,718	6.9	€14.6	build · source
Radeon AI Pro R9700	32	109.73 llama.cpp Vulkan, Q5_K_M (24.68 GB VRAM)	≈€1,718	6.4	€15.7	build · source
Radeon AI Pro R9700	32	102.39 llama.cpp Vulkan, Unsloth UD-Q5_K_XL, short context, f16 KV	≈€1,718	6.0	€16.8	build · source
Radeon AI Pro R9700	32	90.52 llama.cpp Vulkan, UD-Q5_K_XL, 32K context	≈€1,718	5.3	€19.0	build · source
RTX 5090	32	258 llama.cpp, llama-bench tg128 (quant per result)	≈€6,199	4.2	€24.0	build · source
RTX 4090	24	160.6 llama.cpp b8931, Q4_K_M	≈€3,955	4.1	€24.6	build · source
Radeon AI Pro R9700	32	66.43 llama.cpp Vulkan, UD-Q5_K_XL, 128K context	≈€1,718	3.9	€25.9	build · source
Radeon AI Pro R9700	32	44.4 llama.cpp Vulkan, UD-Q5_K_XL, 256K context	≈€1,718	2.6	€38.7	build · source

All measured speeds for Qwen3.6 35B-A3B: best rig for Qwen3.6 35B-A3B.

Qwen3.6 27B: tokens/sec per euro

GPU	VRAM GB	Measured tok/s	GPU price DE (new)	tok/s per €100	€ per tok/s
Radeon AI Pro R9700	32	66.0 llama.cpp Vulkan, Q4_K_M + MTP draft head (draft-n-max 3), 64K context	≈€1,718	3.8	€26.0	build · source
Radeon AI Pro R9700	32	31.9 llama.cpp Vulkan, Q4_K_M, no MTP	≈€1,718	1.9	€53.9	build · source
Radeon AI Pro R9700	32	27.1 llama.cpp HIP (ROCm), Q4_K_M	≈€1,718	1.6	€63.4	build · source
Radeon AI Pro R9700	32	26.7 Ollama (HIP), Q4_K_M	≈€1,718	1.6	€64.3	build · source
RTX 5090	32	75.7 llama.cpp, llama-bench tg128 (quant per result)	≈€6,199	1.2	€81.9	build · source
RTX 5090	32	75.1 llama.cpp 0.7.0-dev, Q4_K_M	≈€6,199	1.2	€82.5	build · source
RTX 4090	24	47.0 llama.cpp, llama-bench tg128 (quant per result)	≈€3,955	1.2	€84.1	build · source

All measured speeds for Qwen3.6 27B: best rig for Qwen3.6 27B.

Qwen3.8 Flash-Next: tokens/sec per euro

GPU	VRAM GB	Measured tok/s	GPU price DE (new)	tok/s per €100	€ per tok/s
2× RTX 5070 Ti	32	182 capy-flash v0.7.0, 2 GPUs, agentic decode, GSQ-RCO (64 GB edition)	≈€2,375	7.7 †	€13.0	build · source
2× RTX 5070 Ti	32	155 capy-flash v0.6.0, 2 GPUs, agentic decode, NVFP4	≈€2,375	6.5 †	€15.3	build · source
2× RTX 5080	32	144.4 capy-flash v0.7.2, --expert-cpu-threads 15, agentic decode, NVFP4	≈€3,198	4.5 †	€22.1	build · source
2× RTX 5080	32	141.3 capy-flash v0.7.2 (CUDA 13.4), agentic decode, NVFP4	≈€3,198	4.4 †	€22.6	build · source
RTX 5090	32	159.2 Strata, IQ3_S (~55 GB), experts offloaded to RAM, speculation helper	≈€6,199	2.6 †	€38.9	build · source
RTX 5090	32	78.8 Strata (296b59d), Unsloth UD-Q5_K_XL, MTP drafting, experts offloaded to RAM	≈€6,199	1.3 †	€78.7	build · source

† Qwen3.8 Flash-Next weights are about 110.1 GB at Q4_K_M / native size, more than this card's VRAM. The source may have used a smaller quant or partial CPU offload; check it before buying.

All measured speeds for Qwen3.8 Flash-Next: best rig for Qwen3.8 Flash-Next.

Gemma 4 12B: tokens/sec per euro

GPU	VRAM GB	Measured tok/s	GPU price DE (new)	tok/s per €100	€ per tok/s
RTX 3060	12	38.5 llama.cpp, llama-bench tg128 (quant per result)	≈€429	9.0	€11.1	build · source
RTX 5070 Ti	16	87.6 llama.cpp, llama-bench tg128 (quant per result)	≈€1,188	7.4	€13.6	build · source
RTX 4090	24	98.1 llama.cpp, llama-bench tg128 (quant per result)	≈€3,955	2.5	€40.3	build · source
RTX 4090	24	97.7 llama.cpp b8931, QAT Q4_0	≈€3,955	2.5	€40.5	build · source
RTX 5090	32	142 llama.cpp, llama-bench tg128 (quant per result)	≈€6,199	2.3	€43.7	build · source

All measured speeds for Gemma 4 12B: best rig for Gemma 4 12B.

Gemma 4 31B: tokens/sec per euro

GPU	VRAM GB	Measured tok/s	GPU price DE (new)	tok/s per €100	€ per tok/s
RTX 5090	32	70.0 llama.cpp, llama-bench tg128 (quant per result)	≈€6,199	1.1	€88.6	build · source
RTX 4090	24	43.8 llama.cpp, llama-bench tg128 (quant per result)	≈€3,955	1.1	€90.3	build · source

All measured speeds for Gemma 4 31B: best rig for Gemma 4 31B.

Gemma 4 26B-A4B: tokens/sec per euro

GPU	VRAM GB	Measured tok/s	GPU price DE (new)	tok/s per €100	€ per tok/s
RTX 4090	24	164 llama.cpp, llama-bench tg128 (quant per result)	≈€3,955	4.1	€24.1	build · source
RTX 5090	32	235 llama.cpp, llama-bench tg128 (quant per result)	≈€6,199	3.8	€26.4	build · source

All measured speeds for Gemma 4 26B-A4B: best rig for Gemma 4 26B-A4B.

Qwen3.5 9B: tokens/sec per euro

GPU	VRAM GB	Measured tok/s	GPU price DE (new)	tok/s per €100	€ per tok/s
RTX 4090	24	137 llama.cpp, llama-bench tg128 (quant per result)	≈€3,955	3.5	€28.9	build · source
RTX 5090	32	208 llama.cpp, llama-bench tg128 (quant per result)	≈€6,199	3.4	€29.8	build · source

All measured speeds for Qwen3.5 9B: best rig for Qwen3.5 9B.

LFM2.5 24B-A2B: tokens/sec per euro

GPU	VRAM GB	Measured tok/s	GPU price DE (new)	tok/s per €100	€ per tok/s
RTX 3060	12	125 llama.cpp, llama-bench tg128 (quant per result)	≈€429	29.1 †	€3.4	build · source
RTX 4090	24	261-339 llama.cpp, llama-bench tg128 (quant per result)	≈€3,955	6.6	€15.2	build · source

† LFM2.5 24B-A2B weights are about 14.7 GB at Q4_K_M / native size, more than this card's VRAM. The source may have used a smaller quant or partial CPU offload; check it before buying.

All measured speeds for LFM2.5 24B-A2B: best rig for LFM2.5 24B-A2B.

Flint 40B-A2.5B: tokens/sec per euro

GPU	VRAM GB	Measured tok/s	GPU price DE (new)	tok/s per €100	€ per tok/s
RTX 4090	24	179.4 llama.cpp 0.7.0-dev (build 11803), NVFP4	≈€3,955	4.5 †	€22.0	build · source
RTX 4090	24	169.2 llama.cpp 0.7.0-dev (build 11803), UD-Q3_K_XL	≈€3,955	4.3 †	€23.4	build · source

† Flint 40B-A2.5B weights are about 24.3 GB at Q4_K_M / native size, more than this card's VRAM. The source may have used a smaller quant or partial CPU offload; check it before buying.

All measured speeds for Flint 40B-A2.5B: best rig for Flint 40B-A2.5B.

Complete systems with a reported price

Laptops, Macs and full builds where the source states what the whole machine cost (converted to € at the ECB rate 1.1208 where needed).

System	Model	tok/s	Reported price	tok/s per €100
DGX Spark (GB10): Qwen3.6 35B-A3B NVFP4 at 113 tok/s on vLLM with MTP	Qwen3.6 35B-A3B	113	$3,999 (≈€3,568)	3.2
Minisforum MS-S1 MAX (Strix Halo, 128 GB): Qwen3.6 35B-A3B at 62 tok/s on llama.cpp Vulkan	Qwen3.6 35B-A3B	93	$3,799 (≈€3,390)	2.7
DGX Spark (GB10): Qwen3.6 35B-A3B NVFP4 at 113 tok/s on vLLM with MTP	Qwen3.6 35B-A3B	86	$3,999 (≈€3,568)	2.4
Minisforum MS-S1 MAX (Strix Halo, 128 GB): Qwen3.6 35B-A3B at 62 tok/s on llama.cpp Vulkan	Qwen3.6 35B-A3B	62	$3,799 (≈€3,390)	1.8
HP OmniBook Ultra 16 (RTX Spark N1X, 128GB) laptop: local AI benchmarks	Qwen3.6 35B-A3B	89.5	$6,230 (≈€5,559)	1.6
Minisforum MS-S1 MAX (Strix Halo, 128 GB): Qwen3.6 35B-A3B at 62 tok/s on llama.cpp Vulkan	Qwen3.6 35B-A3B	54	$3,799 (≈€3,390)	1.6
Minisforum MS-S1 MAX (Strix Halo, 128 GB): Qwen3.6 35B-A3B at 62 tok/s on llama.cpp Vulkan	Qwen3.6 35B-A3B	51	$3,799 (≈€3,390)	1.5
HP OmniBook Ultra 16 (RTX Spark N1X, 128GB) laptop: local AI benchmarks	gpt-oss 20B	58.0	$6,230 (≈€5,559)	1.0
Minisforum MS-S1 MAX (Ryzen AI Max+ PRO 495, 192 GB): gpt-oss 120B at 55.8 tok/s	gpt-oss 120B	55.8	$7,399 (≈€6,602)	0.8
Minisforum MS-S1 MAX (Ryzen AI Max+ PRO 495, 192 GB): gpt-oss 120B at 55.8 tok/s	gpt-oss 120B	45.1	$7,399 (≈€6,602)	0.7
Minisforum MS-S1 MAX (Ryzen AI Max+ PRO 495, 192 GB): gpt-oss 120B at 55.8 tok/s	MiniMax M2.7	29.5	$7,399 (≈€6,602)	0.4
Minisforum MS-S1 MAX (Ryzen AI Max+ PRO 495, 192 GB): gpt-oss 120B at 55.8 tok/s	DeepSeek V4 Flash	17.4	$7,399 (≈€6,602)	0.3
Minisforum MS-S1 MAX (Ryzen AI Max+ PRO 495, 192 GB): gpt-oss 120B at 55.8 tok/s	MiniMax M2.7	17.3	$7,399 (≈€6,602)	0.3
$2K EPYC 64-core, 512 GB DDR4, CPU-only	DeepSeek R1 671B	4.31	$2,000 (≈€1,784)	0.2
Minisforum MS-S1 MAX (Ryzen AI Max+ PRO 495, 192 GB): gpt-oss 120B at 55.8 tok/s	DeepSeek V4 Flash	12.2	$7,399 (≈€6,602)	0.2
Minisforum MS-S1 MAX (Ryzen AI Max+ PRO 495, 192 GB): gpt-oss 120B at 55.8 tok/s	Qwen3.8 27B	8	$7,399 (≈€6,602)	0.1

Not in the tables

RTX 3090: no longer sold new in Germany, so there's no comparable new price. On the used market it's often the value king for 24 GB, see the used 3090 buying guide.
Macs, Strix Halo, DGX Spark, RTX Spark: the GPU can't be bought separately; see Mac vs PC and the system table above.
Cards with no measured 2026-model result in RigBase yet. Submit yours.

Next steps: check the model fits with the VRAM calculator, compare raw speed in best GPU for local LLMs 2026, or browse rigs by budget.

Not enough VRAM? Rent before you buy. Run the model on the exact GPU on a rented 24–96 GB cloud GPU by the hour and see the real speed before spending €1,000+. Try it on Runpod
Referral link: if you sign up with Google and add $10, Runpod gives you and RigBase bonus credit (terms). It never affects our data.

New: Used RTX 3090 vs RTX 5090 · Best GPU under €1,000