Best home rig for Qwen3.8 27B

Will it fit my GPU? Best value GPU per euro →

Real, measured generation speeds for Qwen3.8 27B on home hardware, fastest first. Different quants/engines aren't directly comparable.

Build	Quant	Engine	tok/s	Cost
Dual RTX 5070 Ti on Ryzen 9 9900X (capy-flash maintainer rig)	3-bit GSQ	capy-flash, 2 GPUs, MTP, decode at 8k prompt	230		source
RTX Pro 6000 Blackwell Max-Q (300 W cap): Qwen3.8 27B on vLLM, NVFP4 vs BF16	NVFP4 (W4A16) + MTP (5 draft tokens), 256K context	vLLM 0.27.1	150		source
Single RTX 3090: Qwen3.8 27B at 78-140 tok/s with the OpenJet megakernel vs llama.cpp	Q4_K_M (unsloth), greedy code prompt	OpenJet megakernel (MTP, 4 drafts)	139.7		source
Dual RTX 5070 Ti on Ryzen 9 9900X (capy-flash maintainer rig)	3-bit GSQ	capy-flash (1 GPU), MTP, decode at 8k prompt	139		source
Dual Tesla V100 32 GB on EPYC 7402 (bnk engine)	Q8_0	bnk dense engine, tensor parallel, MTP drafting	104		source
Single Intel Arc Pro B70 with vLLM XPU cookbook	GPTQ Int4 (asym)	vLLM XPU b70 image + MTP, reasoning prompt	96		source
RTX Pro 6000 Blackwell Max-Q (300 W cap): Qwen3.8 27B on vLLM, NVFP4 vs BF16	BF16 + MTP (5 draft tokens), 256K context	vLLM 0.20.1	85		source
Single RTX 3090: Qwen3.8 27B at 78-140 tok/s with the OpenJet megakernel vs llama.cpp	Q4_K_M (unsloth), reasoning mode	OpenJet megakernel (MTP)	78.0		source
RTX 5090 (llm-speed signed runs, Oct 2026): gpt-oss 20B 324 tok/s, Qwen3.8 27B 75 tok/s	Q4_K_M (EPYC 9534 host)	llama.cpp 0.7.0-dev (b11669)	75.2		source
Single RTX 3090: Qwen3.8 27B at 78-140 tok/s with the OpenJet megakernel vs llama.cpp	Q4_K_M (unsloth), greedy code prompt	llama.cpp (MTP, 2 drafts)	72.5		source
Single Intel Arc Pro B70 with vLLM XPU cookbook	GPTQ Int4 (asym)	vLLM XPU b70 image + MTP, prose prompt	66		source
Radeon AI Pro R9700: Qwen3.8 27B on llama.cpp, Ollama and vLLM (Vadi Taslim)	Q4_K_M + MTP speculative decoding, reasoning on	llama.cpp Vulkan	57		source
i7-12700K + RTX 5060 Ti 16GB + RTX 3060 12GB: Qwen3.8 27B from 21 to 51 tok/s (split, MTP, single-GPU)	UD-IQ4_XS on the 5060 Ti + MTP draft on the 3060 (n_max=2)	llama.cpp master (Oct 2026)	51.4		source
Single RTX 3090: Qwen3.8 27B at 78-140 tok/s with the OpenJet megakernel vs llama.cpp	Q4_K_M (unsloth), reasoning mode	llama.cpp (MTP, 2 drafts)	50.9		source
4x Mac Studio M5 Ultra cluster (Thunderbolt 5 RDMA): Muse Spark Open at 36.7 tok/s	BF16, 8,192-token context	MLX, 4-node tensor parallel	50.89		source
Dual Intel Arc Pro B70 (neural.download lab)	FP8	vLLM XPU, TP2, MTP 5	46.01		source
Single RTX 3090: Qwen3.8 27B at 78-140 tok/s with the OpenJet megakernel vs llama.cpp	Q4_K_M (unsloth), no speculation	OpenJet megakernel	42.9		source
Dual Tesla V100 32 GB on EPYC 7402 (bnk engine)	Q8_0	bnk dense engine, tensor parallel, no drafting	42.4		source
Single RTX 3090: Qwen3.8 27B at 78-140 tok/s with the OpenJet megakernel vs llama.cpp	Q4_K_M (unsloth), no speculation	llama.cpp	40.4		source
RTX 3090: 2026 models (llamabench community)	llama-bench tg128 (quant per result)	llama.cpp	38.2-71.6		source
Mac Studio M5 Max 36 GB: Qwen3.6 35B-A3B at 173 tok/s (Rapid-MLX), 146 on oMLX	4-bit MLX	Rapid-MLX 0.16.1 (MTP)	37.9		source
i7-12700K + RTX 5060 Ti 16GB + RTX 3060 12GB: Qwen3.8 27B from 21 to 51 tok/s (split, MTP, single-GPU)	Q4_K_M, split + MTP draft (n_max=2)	llama.cpp master (Oct 2026)	37.6		source
2x Mac Studio M5 Ultra cluster (Thunderbolt 5 RDMA): Muse Spark Open 816B	BF16, 8,192-token context	MLX, 2-node tensor parallel	33.27		source
MacBook Pro M5 Max 128 GB: Qwen3.6 35B-A3B at 271 tok/s (mlx-serve + MTP), Gemma 4 26B-A4B at 166	4-bit MLX	same server, no speculation	31		source
Radeon AI Pro R9700: Qwen3.8 27B on llama.cpp, Ollama and vLLM (Vadi Taslim)	Q4_K_M	Ollama	30		source
i7-12700K + RTX 5060 Ti 16GB + RTX 3060 12GB: Qwen3.8 27B from 21 to 51 tok/s (split, MTP, single-GPU)	UD-IQ4_XS on the 5060 Ti alone, 8K ctx, q8 KV	llama.cpp master (Oct 2026)	28.0		source
Single Intel Arc Pro B70 32 GB (neural.download lab)	Q4_K_M	llama.cpp SYCL + neural.download TP1 stack, no speculation	27.8		source
Mac Studio M5 Max 36 GB: Qwen3.6 35B-A3B at 173 tok/s (Rapid-MLX), 146 on oMLX	4-bit MLX	LM Studio (MLX)	26.0		source
Mac Studio M5 Max 36 GB: Qwen3.6 35B-A3B at 173 tok/s (Rapid-MLX), 146 on oMLX	4-bit MLX	oMLX 0.7.0	25.3		source
Radeon AI Pro R9700: Qwen3.8 27B on llama.cpp, Ollama and vLLM (Vadi Taslim)	AWQ INT4, single request	vLLM (radiance)	22		source
i7-12700K + RTX 5060 Ti 16GB + RTX 3060 12GB: Qwen3.8 27B from 21 to 51 tok/s (split, MTP, single-GPU)	Q4_K_M, layer split 5/3 across both GPUs	llama.cpp (May 2026 build)	21.2		source
Mac Studio M5 Ultra 512GB: Qwen3.8 27B and Mistral Medium 3.5 (Apple MLX benchmark)	BF16, 8,192-token context	MLX (mlx_lm.benchmark)	19.69		source
DGX Spark (GB10, 128 GB): 23 models on vLLM at 128K context, Qwen3.6 35B-A3B up to 102 tok/s	+ MTP (k=2)	vLLM 0.31.0, 128K ctx, 1 user	19.0		source
Strix Halo (Ryzen AI Max+ 395, 128 GB): Qwen3.8 27B Q6_K, 9.6 to 18 tok/s with MTP	Q6_K + MTP (draft-n-max 2)	llama.cpp Vulkan (RADV)	18.01		source
Strix Halo (Corsair AI Workstation 300, 128 GB): ROCm vs Vulkan on Qwen3.8 27B and Qwen3.6 35B-A3B	Q8_0 + MTP (draft-n-max 3), coding workload, reasoning on	llama.cpp ROCm	15.67		source
Strix Halo (Ryzen AI Max+ 395, 128 GB): Qwen3.8 27B Q6_K, 9.6 to 18 tok/s with MTP	Q6_K + MTP (draft-n-max 5)	llama.cpp Vulkan (RADV)	15.16		source
Microsoft Surface Laptop Ultra (RTX Spark N1X): llama.cpp speeds on 2026 models (Micro Center)	general, short prompt	llama.cpp (Nvidia ARM64 build)	14.5		source
Strix Halo (Corsair AI Workstation 300, 128 GB): ROCm vs Vulkan on Qwen3.8 27B and Qwen3.6 35B-A3B	Q8_0 + MTP (draft-n-max 3), coding workload, reasoning on	llama.cpp Vulkan	13.19		source
Strix Halo mini PC (Ryzen AI Max+ 395, 128 GB): Qwen3.8 27B at 12.6 tok/s Q4_K_M, 7.7 Q8_0	Q4_K_M	llama.cpp/Ollama, ROCm	12.6		source
Mac Studio M3 Ultra 512GB: Qwen3.8 27B and Mistral Medium 3.5 (Apple MLX benchmark)	BF16, 8,192-token context	MLX (mlx_lm.benchmark)	12.46		source
DGX Spark (GB10, 128 GB): 23 models on vLLM at 128K context, Qwen3.6 35B-A3B up to 102 tok/s	no speculative decoding	vLLM 0.31.0, 128K ctx, 1 user	11.0		source
RTX 4060 Ti 16GB: 2026 models (llamabench community)	llama-bench tg128 (quant per result)	llama.cpp	10.7-17.6		source
Strix Halo (Ryzen AI Max+ 395, 128 GB): Qwen3.8 27B Q6_K, 9.6 to 18 tok/s with MTP	Q6_K, no MTP	llama.cpp Vulkan (RADV)	9.59		source
Minisforum MS-S1 MAX (Ryzen AI Max+ PRO 495, 192 GB): gpt-oss 120B at 55.8 tok/s	Q8 (author: "about 8 tok/s")	llama.cpp Vulkan	8	$7,399 (≈€6,602)	source
Strix Halo (Corsair AI Workstation 300, 128 GB): ROCm vs Vulkan on Qwen3.8 27B and Qwen3.6 35B-A3B	Q8_0, llama-bench tg128	llama.cpp Vulkan	7.77		source
Strix Halo (Corsair AI Workstation 300, 128 GB): ROCm vs Vulkan on Qwen3.8 27B and Qwen3.6 35B-A3B	Q8_0, llama-bench tg128	llama.cpp ROCm	7.76		source
Strix Halo mini PC (Ryzen AI Max+ 395, 128 GB): Qwen3.8 27B at 12.6 tok/s Q4_K_M, 7.7 Q8_0	Q8_0, --reasoning off	llama.cpp/Ollama, ROCm	7.7		source
RTX 3060 12GB: 2026 models (llamabench community)	llama-bench tg128 (quant per result)	llama.cpp	5.2-6.7		source

Rent before you buy: try this model on a cloud GPU for a few dollars first: Runpod (referral link; sign up with Google to get the new-user bonus)

Compare: DGX Spark vs Mac Studio vs Strix Halo: measured tok/s and tok/s per watt

Memory math: how much VRAM / RAM does Qwen3.8 27B need?

Qwen3.8 27B has 27B parameters (source: Ollama library tag 27b (nominal)). Weight size = parameters × bits-per-weight ÷ 8, before context (KV cache) and runtime overhead.

Format	Bits/weight	Weights	Smallest common memory size that holds the weights alone
Q8_0	8.50	28.7 GB	32 GB
Q6_K	6.56	22.2 GB	24 GB
Q5_K_M	5.70	19.2 GB	24 GB
Q4_K_M	4.89	16.5 GB	24 GB
Q3_K_M	4.00	13.5 GB	16 GB
Q2_K	3.16	10.7 GB	12 GB

Bits-per-weight for GGUF quants are llama.cpp's own figures (llama.cpp quantize README). "As measured" rows apply the source's stated bits-per-weight to the parameter count; real files differ slightly because some layers stay at higher precision. Try other combinations in the VRAM calculator.

Measured on setups with less GPU memory than the Q4_K_M weights (so part of the model lived in system RAM or was streamed from disk):

RTX 3060 (12 GB): 5.2-6.7 tok/s (llama-bench tg128 (quant per result))
RTX 4060 Ti 16GB (16 GB): 10.7-17.6 tok/s (llama-bench tg128 (quant per result))

FAQ

How much VRAM does Qwen3.8 27B need?

The weights are about 16.5 GB at Q4_K_M (27B params × 4.89 bits ÷ 8, before context/KV cache). Add memory for context (KV cache) and the runtime. The smallest-memory setup with a measured result here is RTX 5070 Ti (16 GB) at 139 tok/s (3-bit GSQ). At Q4_K_M the weights are about 16.5 GB, more than 16 GB, so Q4_K_M doesn't fully fit without offloading part of the model to system RAM.

What is the fastest measured hardware for Qwen3.8 27B?

2x RTX 5070 Ti: 230 tok/s generation (capy-flash, 2 GPUs, MTP, decode at 8k prompt, 3-bit GSQ), from https://github.com/ValerioDolci/capy-flash.

Which GPUs have measured results for Qwen3.8 27B?

2x Apple M5 Ultra, 2x Intel Arc Pro B70, 2x RTX 5070 Ti, 2x Tesla V100 32GB, 4x Apple M5 Ultra, Apple M3 Ultra, Apple M5 Ultra, DGX Spark (GB10), Intel Arc Pro B70, RTX 3060 12GB, RTX 3090, RTX 4060 Ti 16GB, RTX 5070 Ti, RTX 5090, RTX Pro 6000 Max-Q (Blackwell, 96 GB, 300 W), RTX Spark N1X, Radeon AI Pro R9700, Strix Halo (Radeon 8060S).

Can it run on… 2x Apple M5 Ultra · 2x Intel Arc Pro B70 · 2x RTX 5070 Ti · 2x Tesla V100 32GB · 4x Apple M5 Ultra · Apple M3 Ultra · Apple M5 Ultra · DGX Spark (GB10) · Intel Arc Pro B70 · RTX 3060 12GB · RTX 3090 · RTX 4060 Ti 16GB · RTX 5070 Ti · RTX 5090 · RTX Pro 6000 Max-Q (Blackwell, 96 GB, 300 W) · RTX Spark N1X · Radeon AI Pro R9700 · Strix Halo (Radeon 8060S)