24 GB value workstation
Practical local entry point24 GB VRAM. A pragmatic used-GPU build for quantized local models without workstation pricing. $1,500.00 whole-system planning allowance.
Planning assumption ↗The short catalog
Every option should earn its slot.
The catalog stays small and opinionated. Each entry links its evidence, so when a recommendation looks wrong to you, the receipts are one click away.
Open the planner6 compared
Hardware classes
Planning envelopes for the machines the engine can recommend.
Mac mini (M4 Pro, 24 GB)
Turnkey Apple Silicon entry point24 GB unified · 273 GB/s. A quiet, turnkey complete system at list price for MLX or llama.cpp local inference. $1,399.00 first-party list price.
Apple Mac mini newsroom ↗Mac Studio (M4 Max, 36 GB)
Unified-memory local workstation36 GB unified · 410 GB/s. A quiet, turnkey complete system at list price with more unified memory for MLX or llama.cpp serving. $1,999.00 first-party list price.
Apple Mac Studio newsroom ↗32 GB performance workstation
Fast local default32 GB VRAM · 1,792 GB/s. Fast local inference for everyday models, backed by APIs only when the job merits it. $3,200.00 whole-system planning allowance.
NVIDIA RTX 5090 specifications ↗Mac Studio (M3 Ultra, 96 GB)
High-capacity Apple Silicon workstation96 GB unified · 819 GB/s. A quiet, turnkey complete system at list price for capacity-first MLX or llama.cpp serving. $3,999.00 first-party list price.
Apple Mac Studio newsroom ↗DGX Spark capacity appliance
Local model capacity128 GB unified · 273 GB/s. Prioritizes model capacity and privacy over the raw token speed of a consumer GPU. $4,699.00 first-party list price.
NVIDIA DGX Spark hardware ↗13 compared
Hosted inference
Elastic routes for bursty workloads and the jobs that exceed local capability.
DeepSeek V4 Flash 0731
Value coding and batch route$0.14 input / $0.28 output per million tokens; uncached pricing.
DeepSeek API pricing ↗GPT-5.4 nano
Lowest-cost GPT text route$0.20 input / $1.25 output per million tokens.
OpenAI API pricing ↗Gemini 3.5 Flash-Lite
Low-cost multimodal route$0.30 input / $2.50 output per million tokens.
Gemini Developer API pricing ↗Claude Haiku 4.5
Focused Claude economy route$1.00 input / $5.00 output per million tokens.
Claude API pricing ↗GPT-5.6 Luna
Research and generalist route$0.20 input / $1.20 output per million tokens.
OpenAI API pricing ↗DeepSeek V4 Pro
Stronger DeepSeek coding route$0.44 input / $0.87 output per million tokens; uncached pricing.
DeepSeek API pricing ↗GPT-5.4 mini
Compact GPT generalist$0.75 input / $4.50 output per million tokens.
OpenAI API pricing ↗Gemini 3.6 Flash
Fast multimodal generalist$1.50 input / $7.50 output per million tokens.
Gemini Developer API pricing ↗Claude Sonnet 5
Premium Claude generalist$2.00 input / $10.00 output per million tokens; introductory rate through August 31, 2026.
Claude API pricing ↗GPT-5.6 Terra
Mid-tier GPT-5.6 generalist$2.00 input / $12.00 output per million tokens.
OpenAI API pricing ↗Gemini 3.1 Pro
Premium Gemini multimodal route$2.00 input / $12.00 output per million tokens; paid tier for prompts up to 200K tokens.
Gemini Developer API pricing ↗GPT-5.6 Sol
Difficult-task escalation$5.00 input / $30.00 output per million tokens; never used as the bulk route.
OpenAI API pricing ↗Claude Opus 5
Premium difficult-task escalation$5.00 input / $25.00 output per million tokens; never used as the bulk route.
Claude API pricing ↗8 compared
Local models
Private utility and reasoning routes sized to hardware a single developer could own.
Gemma 4 E4B
Existing-machine utility modelA small multimodal utility model for light local work on hardware you already have.
Google Gemma 4 overview ↗gpt-oss-20b
Entry-point coding MoEA compact local reasoning model for coding and document work.
OpenAI gpt-oss-20b model card ↗Mistral Small 3.3
24 GB dense multimodal generalistA compact local multimodal model for coding and document work.
Mistral Small 3.3 model card ↗Gemma 4 26B A4B (quantized)
Consumer-GPU generalistThe local multimodal generalist for 24–32 GB consumer GPUs.
Google Gemma 4 getting started ↗Qwen3.6 27B
24 GB dense coding generalistA dense local model for coding and research on 24 GB class hardware.
Qwen3.6 27B model card ↗Llama 4 Scout
Capacity-class multimodal MoEA large local multimodal model for private research and creative work.
Meta Llama 4 Scout model card ↗gpt-oss-120b
Capacity-class coding MoE for 80 GB+ machinesA large local reasoning model for technical work that fits on an 80 GB GPU.
OpenAI gpt-oss-120b model card ↗DeepSeek V4 Flash 0731 (experimental Q2)
Large local coding routeA large local coding and reasoning option; viable only with aggressive quantization and careful serving.
Official DeepSeek model weights ↗Quick starts
Inspect a complete preset.
Solo developer defaultA balanced hybrid setup for daily coding, research, and document work.View stack →API-first value stackKeep current hardware and route light work to inexpensive hosted models.View stack →Private local labA capacity-first one-person setup for local-only coding, vision, and documents.View stack →