Only with CPU offload
GLM-4.7-Flash 30B-A3B at Q4_K_M needs 18.6 GB but only 10.6 GB is addressable, so about 45% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 15 tokens per second — usable for batch work, painful for chat.
55% on GPU
8K context
Q4_K_M · 17.5 GB
MIT
Released 20 Jan 2026
MIT-licensed 30B-A3B tuned for agentic coding, with a DeepSeek-style latent KV cache. 60–80 tok/s reported on a 4090.
The VRAM budget
weights 17.5 GB
Weights 17.5 GB
KV cache @ 8K 0.41 GB
Runtime overhead 0.6 GB
Over budget 8.0 GB past 10.6 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 30.9 GB | 31.9 GB | — | ~6.3 | −0.1% ppl | 21.3 GB over |
| Q6_K | 23.8 GB | 24.8 GB | — | ~9.1 | −0.4% ppl | 14.2 GB over |
| Q5_K_M | 20.6 GB | 21.6 GB | — | ~11 | −0.8% ppl | 11.0 GB over |
| Q4_K_M | 17.5 GB | 18.6 GB | — | ~15 | −1.9% ppl | 8.0 GB over |
| Q3_K_M | 14.2 GB | 15.2 GB | — | ~23 | −5.4% ppl | 4.6 GB over |
| Q2_K | 12.2 GB | 13.2 GB | — | ~34 | −15% ppl | 2.6 GB over |
Quality is the published perplexity delta against f16 weights. Max context assumes an f16 KV cache; q8_0 roughly doubles it. This model uses multi-head latent attention, so its cache is a compressed latent rather than full K and V.
How to run it
$ llama-server \
-hf zai-org/GLM-4.7-Flash:Q4_K_M \
-c 8192 -ngl 25
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.
01Download is 17.5 GB. Keep it on an SSD — a first load off a spinning disk takes minutes.
02Close anything else holding VRAM. A browser with hardware acceleration can sit on 1–2 GB.
03If it falls back to CPU silently, drop the context first, then step down a quantisation.