LLM hardware buying guide

Choosing a Mac? Inspect current MLX downloads and runtime requirements before deciding how much memory to buy.

• Best hardware for a local coding agent

  Pick a rig that runs Qwen3-Coder-Next, Qwen2.5-Coder-32B, gpt-oss, and DeepSeek as a daily-driver coding agent without you waiting on it.

• Best rig for Qwen3-Coder-Next

  Qwen3-Coder-Next is an 80B-parameter MoE with ~3B active. The activation pattern means it punches above its weight on Apple Silicon and sane consumer GPUs. Here's the data we have.

• Fastest rig for Qwen2.5-Coder-32B (local)

  Qwen2.5-Coder-32B is the strongest fully-open dense coding model that fits a single prosumer card at 4-bit. Here's the fastest decode tok/s submitted across every GPU and Apple Silicon tier.

• Fastest rig for Qwen3-Coder-30B-A3B (local)

  Qwen3-Coder-30B-A3B is a 30B mixture-of-experts with ~3B active — it decodes like a small model but codes like a big one. Here's the fastest decode tok/s submitted across every GPU and Apple Silicon tier.

• Local vs hosted: when does buying a GPU pay off?

  At low usage, hosted APIs win on $/Mtok. At high sustained usage, a 4090 or M3 Ultra wins. Here's the break-even math, run against live numbers.

• Best Mac for running local LLMs

  Choose a Mac for your model and workload: current MLX downloads, memory and runtime checks, plus recent submitted short-chat examples with source settings.

• Cheapest GPU for Local LLMs: Measured RTX 3090 vs 4090

  Measured: Qwen3.8-27B Q4_K_M decodes 37.65 tok/s on an RTX 3090 and 43.06 on an RTX 4090 (llama.cpp, 24GB each). Which to buy under $2,000.

• Best GPU for Local LLMs (2026): Measured tok/s on RTX 5090, 4090, 3090

  Measured decode tok/s for RTX 5090, 4090 and 3090 across 8B to 32B models, plus what fits in 24GB vs 32GB. Every number is from a signed benchmark run.

• Cheapest 70B LLM Rig? Measured Options Compared

  Llama-3.3-70B Q4_K_M, rented: two RTX 3090s ran 47% faster (19.1 vs 13.0 tok/s); the A40 cost less per hour and per token at Oct 2026 rates. Runs linked.

• Llama 3.3 70B on a GPU: Speed and VRAM Fit

  At 4-bit, Llama 3.3 70B needs ~48GB. Measured: two RTX 3090s ran it at 19.1 tok/s, an A40 at 13.0. Only IQ1/IQ2 quants fit 24GB. GGUF fit table for 24-80GB.

• Apple Silicon vs NVIDIA for Local LLMs: Measured Speed vs Memory

  Measured: a 96GB M3 Ultra ran Llama-3.3-70B at 16.8 tok/s (MLX); an RTX 5090 ran gpt-oss-20b at 318.4 tok/s (llama.cpp). Different runtimes, not matched.

• Fastest small-MoE coder on a Mac (Qwen3-Coder-30B, DeepSeek-Coder-V2-Lite)

  Two MoE coders — Qwen3-Coder-30B-A3B and DeepSeek-Coder-V2-Lite — both clear 100 tok/s decode on an M3 Ultra at 4-bit. Here's how they rank across every Apple Silicon tier we have data for.