Recipes · 11 configurations
Machines ready to copy For every budget, a model, the hardware, the engine and the command to run. Memory, speed and price use the same formulas as the rest of the site: estimates, not measurements.
Budget €1,180
Your first local LLM on a budget The cheapest way in to chat, summarise or code with a recent model, with no subscription.
Model Qwen3 8BHardware RTX 3060 12 GBPrecision Q4_K_MEngine OllamaContext 8k
Memory 7 / 12 GB
Estimated speed ~39 tok/s
Power 170 W
Memory use 56 %
ollama run qwen3:8bCopy
Budget €1,650
Best price-to-speed in 24 GB A used RTX 3090 and a MoE model: only 3 billion parameters work on each token, hence very high speed.
Model Qwen3 30B-A3BHardware RTX 3090 24 GB (used)Precision Q4_K_MEngine llama.cppContext 16k
Memory 21 / 24 GB
Estimated speed ~242 tok/s
Power 350 W
Memory use 88 %
llama-server -hf unsloth/Qwen3-30B-A3B-GGUF:Q4_K_M -ngl 99 -c 16384Copy
Budget €2,400
Llama 70B for under €2,500 Two used 3090s share the model’s layers. The classic build of local-AI enthusiasts.
Model Llama 3.3 70BHardware 2 × RTX 3090 24 GB (used)Precision Q4_K_MEngine llama.cppContext 8k
Memory 47 / 48 GB
Estimated speed ~12 tok/s
Power 700 W
Memory use 99 %
llama-server -hf unsloth/Llama-3.3-70B-Instruct-GGUF:Q4_K_M -ngl 99 -c 8192Copy
Enthusiast €3,500
A 32B model on a single card The RTX 5090’s 32 GB fit a 32B in Q4 with a long context, and its bandwidth makes it very responsive.
Model Qwen3 32BHardware RTX 5090 32 GBPrecision Q4_K_MEngine llama.cppContext 32k
Memory 29 / 32 GB
Estimated speed ~50 tok/s
Power 575 W
Memory use 91 %
llama-server -hf unsloth/Qwen3-32B-GGUF:Q4_K_M -ngl 99 -c 32768Copy
Compact €2,300
A big MoE in a quiet mini-PC 128 GB of unified memory for €2,300: ideal for MoE models, which stay fast despite modest bandwidth.
Model Qwen3 30B-A3BHardware Mini-PC Ryzen AI Max+ 395 128 GBPrecision Q8_0Engine llama.cppContext 16k
Memory 36 / 109 GB
Estimated speed ~39 tok/s
Power 140 W
Memory use 33 %
llama-server -hf unsloth/Qwen3-30B-A3B-GGUF:Q8_0 -ngl 99 -c 16384Copy
Compact €4,000
gpt-oss-120b in a desktop box OpenAI’s open model fits entirely in the DGX Spark, with room left for a long context.
Model gpt-oss-120bHardware NVIDIA DGX Spark 128 GBPrecision Q4_K_MEngine OllamaContext 32k
Memory 76 / 109 GB
Estimated speed ~47 tok/s
Power 240 W
Memory use 70 %
ollama run gpt-oss:120bCopy
Compact €4,400
Llama 70B on a Mac Studio Silent, efficient and ready to use. Slower than a pair of GPUs, but nothing to assemble.
Model Llama 3.3 70BHardware Mac Studio M4 Max 128 GBPrecision Q4_K_MEngine OllamaContext 16k
Memory 50 / 96 GB
Estimated speed ~7.1 tok/s
Power 150 W
Memory use 52 %
ollama run llama3.3:70bCopy
Compact €11,500
Qwen3 235B on a single Mac 512 GB of unified memory: the only desktop machine that loads a model this size without a server.
Model Qwen3 235B-A22BHardware Mac Studio M3 Ultra 512 GBPrecision Q4_K_MEngine OllamaContext 32k
Memory 154 / 384 GB
Estimated speed ~34 tok/s
Power 270 W
Memory use 40 %
ollama run qwen3:235bCopy
Pro €9,900
gpt-oss-120b on a workstation 96 GB on one card: a 120-billion-parameter model served very fast, without multi-GPU.
Model gpt-oss-120bHardware RTX PRO 6000 Blackwell 96 GBPrecision Q4_K_MEngine vLLMContext 32k
Memory 76 / 96 GB
Estimated speed ~307 tok/s
Power 600 W
Memory use 80 %
vllm serve openai/gpt-oss-120bCopy
Team €36,000
Team server: 70B for 10 people Four L40S in FP8: quality almost intact and enough cache for ten parallel conversations.
Model Llama 3.3 70BHardware 4 × L40S 48 GBPrecision FP8Engine vLLMContext 8k × 10 users
Memory 99 / 192 GB
Estimated speed ~20 tok/s
Power 1,400 W
Memory use 52 %
vllm serve meta-llama/Llama-3.3-70B-Instruct --quantization fp8 --tensor-parallel-size 4 --max-model-len 8192Copy
Datacenter €270,000
Full DeepSeek R1 in FP8 The full reasoning model, uncompressed, on a node of 8 H200s (1.1 TB of HBM memory).
Model DeepSeek R1 671BHardware 8 × H200 141 GBPrecision FP8Engine vLLMContext 32k × 4 users
Memory 714 / 1,128 GB
Estimated speed ~426 tok/s
Power 5,600 W
Memory use 63 %
vllm serve deepseek-ai/DeepSeek-R1 --tensor-parallel-size 8Copy