Skip to content
llm-speed

Qwen3-Coder 30B speed: 5090, M4 Max, M3 Ultra

Published 2026-07-02

Qwen3-Coder-30B-A3B is one of the strongest open coding models you can run locally: a 30B mixture-of-experts that keeps only about 3B parameters active per token. That design makes it fast on a surprising range of hardware. Here is what community-submitted runs (not measured by llm-speed) recorded, signed and reproducible, each platform on the backend it runs best. Every number links to its run.

Decode speed by hardware

RTX 5090: 260 tok/s. RTX 4090: 145.2 tok/s. Apple M4 Max: 113 tok/s. Apple M3 Ultra: 112 tok/s. All four are the single-request chat-short workload. The RTX 5090 leads, the 4090 is about 1.8x behind it, and both Apple chips land near 112. (The same 4090 run reached 179.9 tok/s on the separate concurrent-decode workload, which is not comparable with these.) Full, current standings are on the cheatsheet and the model page.

Why it is quick even on a laptop-class chip

Decode speed tracks the memory traffic of the parameters that are active each token, not the model's total size. With only ~3B active, Qwen3-Coder moves far less data per token than a dense 30B, so even an Apple chip clears 100 tok/s (roughly 4x reading speed). The same effect is why small-MoE coders beat dense 32B models on the same card.

What should you run it on?

It fits comfortably at Q4 on any 24 GB or larger GPU, and on a maxed Apple. For maximum speed pick the RTX 5090; the RTX 4090 is close behind and cheaper; an Apple M-series runs it quietly at about 112. Confirm the fit for your card with the VRAM-fit checker.

Is that fast enough for a coding agent?

Comfortably. Anything above roughly 40 tok/s already outpaces reading speed, so even the Apple result near 112 tok/s streams edits faster than you can follow, and the 4090 at 145 keeps up with long agent traces that emit thousands of tokens per turn. For a model this capable, the bottleneck is rarely the decode rate. The tok/s thresholds are on the glossary.

Reproduce it, or add your rig

The same one-line install measures any of these:

$ pipx install llm-speed && llm-speed bench

Numbers as of July 2026; the linked runs and the cheatsheet always reflect current data.