unfuse-GLM-4.6-Flash-text-GGUF

GLM-4.6-Flash-text GGUF

GGUF quantizations of sartajbhuvaji/unfuse-GLM-4.6-Flash-text, which is zai-org/GLM-4.6V-Flash with its 892M-parameter vision stack removed.

9,400,279,040 params, glm4 architecture, 131072 context, text-only.

Sizes are GB (10⁹ bytes) as the Hub reports them. ls -h shows smaller GiB numbers for the same files.

Usage

llama-cli -hf sartajbhuvaji/unfuse-GLM-4.6-Flash-text-GGUF:Q4_K_M \
  -p "Explain gradient descent" -n 400 -st

ollama run hf.co/sartajbhuvaji/unfuse-GLM-4.6-Flash-text-GGUF:Q4_K_M

-st (--single-turn) matters for scripted use. Without it llama-cli drops into interactive mode and waits on stdin. The older -no-cnv flag has been removed from current llama.cpp.

Budget your token limit. This is a reasoning model: it emits a thinking block before answering, sometimes in Chinese regardless of prompt language. 160 tokens is not enough to get past the reasoning to an answer, so use 600 or more.

What was removed

The parent model is Glm4vForConditionalGeneration: a 24-layer ViT feeding soft tokens into a GLM-4 decoder via masked_scatter. Text tokens never touch a vision weight, so deleting the branch leaves the text computation alone.

892,498,432 params removed, 8.671% of the model.

The extraction was verified bit-exact against the original on text input: max|d| = 0.000e+00 across six prompts and again at 1,207 tokens. Full detail and the architecture diagram are on the parent model card.

Verification of these quants

Each file was checked to load and generate coherent text under llama.cpp (architecture: glm4 recognised, correct param count and context length). No perplexity or benchmark comparison against bf16 was run. The bit-exactness result above applies to the bf16 weights, not to these lossy quantizations. If you need a measured quality delta, compute it yourself.

Provenance

Quantized with llama.cpp at master from the bf16 weights in the parent repo. Derived from zai-org/GLM-4.6V-Flash (MIT). This repo is MIT as well.