How fast is GLM-4.7-Flash on an RTX 4090?

GLM-4.7-Flash is a small mixture-of-experts coding model: about 30B total parameters with roughly 3B active per token, and a long context window. The low active-parameter count is the whole point. It keeps decode fast while the model still fits a single 24GB card at 4-bit. So how fast does it actually run? We measured it, signed and reproducible.

Decode speed on an RTX 4090

On a 24GB RTX 4090, GLM-4.7-Flash decodes 130 tok/s at Q4, with room left for context. That is well above reading speed and comfortable for an interactive coding assistant on one consumer card. Because only about 3B parameters fire per token, the decode rate tracks that small active slice, not the full 30B.

How it compares to other small-MoE coders

GLM-4.7-Flash sits in the same class as Qwen3-Coder-30B-A3B, which decodes 180 tok/s on the same 4090. Both are 30B-class MoE coders with a few billion active parameters, both fit 24GB, and both run several times faster than a dense model of similar total size. Pick by coding quality and context length rather than by which name you recognize. See the fastest local coding models for the wider field.

Reproduce it, or add your rig

The same one-line install measures it on your own hardware:

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

Number measured July 2026 on a 24GB RTX 4090 at Q4; the linked run and the cheatsheet always reflect current data.