The real cost to run Qwen3 4B Instruct 2507 NVFP4

llmat · 2.8B parameters · 262.1K context · commercial OK

Qwen3 4B Instruct 2507 NVFP4 — 2.8B params (llmat). Auto-indexed from the Hugging Face Hub (116,096 downloads). Parameter count is exact; download size and quantizations are estimates.

What it costs to run Qwen3 4B Instruct 2507 NVFP4 — $0 markup

• On your own machine — $0. Runs free locally if you have about 5 GB of VRAM at Q4_K_M (point Spanvero at LM Studio, Ollama, or llama.cpp).
• On your own rented GPU — from $0.26/hr. NVIDIA RTX 3090 24GB at the direct Vast.ai price (≈ $0.04 a session), $0 markup.
• Via your own API key — $0.12/1M tokens (blended input + output). No GPU to manage (rough estimate for this size).

When does owning hardware beat the API?

At $0.12/1M tokens, the cheapest API for Qwen3 4B Instruct 2507 NVFP4 stays cheaper than a used RTX 3090 at any daily volume — electricity alone costs more per token than the API does.

Local is never $0 — every figure here includes electricity and hardware amortization. Speeds come from published 7-8B Q4 benchmarks (your model, quant, and context will differ); break-even volumes are rounded to 2 significant figures.

Key facts

Open the free Spanvero advisor → for the live, interactive math for your exact workload and hardware.

Related models

• Qwen3 0.6B — 800M, Qwen
• Qwen2.5 3B Instruct — 3.1B, Qwen
• Llama 3.2 3B Instruct — 3B, Meta
• Qwen2.5 1.5B Instruct — 1.5B, Qwen
• gemma 3 270m — 300M, google

Browse: All models · Compare

Spanvero · What's new · Prices as of 2026-07-06. We're an honest advisor — $0 markup, your own accounts, we never resell compute. © 2026 Cynosure LLC.