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.
For a 30B-class coding model, the fastest measured setup is an RTX 5090: Qwen3-Coder-30B-A3B decoded at 259.9 tok/s (llama.cpp, r_c7qyvvmmsv1), vs 145.2 tok/s on a 24GB RTX 4090 (Ollama Q4_K_M, r_wjq32z47vlp), 112.2 tok/s on a 96GB M3 Ultra (MLX, r_fpsca03u2o_) and 113.3 tok/s on an M4 Max 128GB (Ollama, r_roktphpc--8). On a dense 32B coder, the 5090 measured 71.9 tok/s for Qwen2.5-Coder-32B (r_nkbs6d3-d21). These are short-chat speeds; we haven't measured coding quality or long agent prompts at full context.
Fastest coder-model run on the leaderboard: DeepSeek-Coder-V2-Lite-Instruct on RTX 5090 (32GB) at 309.5tok/s.
A local coding agent stresses three things at once: enough VRAM (or unified memory) to hold a 14B-32B coder model at a reasonable quantization, prefill speed for long tool-use prompts, and decode speed for the model's reply. Below are the configurations we have benchmark data for, ranked by decode tok/s on the workloads tagged 'coder'. We list every result so you can make the call yourself, not just the headline number.
Submitted benchmarks
| Hardware | Coder model | decode tok/s | Run |
|---|---|---|---|
| RTX 5090 (32GB) | DeepSeek-Coder-V2-Lite-Instruct | 309.5tok/s | r_0_gs1rgl2fl |
| RTX 4090 (24GB) | Qwen3-Coder-30B-A3B-Instruct | 232.5tok/s | r_1k8942tw3zf |
| RTX 3090 (24GB) | Qwen3-Coder-30B-A3B-Instruct | 199.7tok/s | r_0dh1kzbv4h0 |
| M3 Ultra (60-core GPU) | mlx-community/DeepSeek-Coder-V2-Lite-Instruct-4bit | 168.3tok/s | r_l_v1-zq_qaz |
| M4 Max (40-core GPU) | qwen3-coder | 109.9tok/s | r_r0di2hkku1h |
See also: All hardware · All models · Methodology