GLM-5.3-Flash-UNCENSORED — R2-40 IQ5_KS expert GGUF (for flash3090)

Routed-expert weights for running GLM-5.3-Flash (320B total / 18B active MoE) on one or two RTX 3090s with the flash3090 engine. flash3090 reads the routed experts from this file (CPU + VRAM hot cache) and the dense backbone from the FP8 checkpoint dealignai/GLM-5.3-Flash-UNCENSORED-FP8.

Base: zai-org/GLM-5.3-Flash (MIT), uncensored derivative by dealignai (MIT). This is an uncensored model: safety refusals were removed by the derivative's author. Use responsibly.
Quantization: ik_llama.cpp llama-quantize, routed experts IQ5_KS (last MoE layer IQ6_K), shared experts / dense FFN Q8_0, attention and indexer BF16; ~207 GB in 13 shards; importance matrix by AesSedai. Full recipe:

--custom-q 'blk\.45\.ffn_(gate|up|down)_exps\.weight=iq6_k,ffn_(gate|up|down)_exps\.weight=iq5_ks,ffn_(gate|up|down)_shexp\.weight=q8_0,ffn_(gate|up|down)\.weight=q8_0,indexer\.(attn_k|attn_q_b|proj)\.weight|indexer_compressor_(gate|ape)\.weight=bf16,hc_(attn|ffn)_fn\.weight=f32,attn_(k_b|v_b)\.weight=bf16,attn_(q_a|q_b)\.weight=q8_0,attn_kv_a_mqa\.weight=bf16,attn_(q|k|v|output)\.weight=bf16,ssm_(f_a|f_b|g_a|g_b|beta)\.weight=bf16,nextn\.eh_proj\.weight=bf16,token_embd\.weight=q8_0,output\.weight=bf16'

Quality: mean KL divergence vs BF16 of 0.026–0.036 on held-out text, on par with ik_llama.cpp on the same file.
Also included: mmproj-GLM-5.3-Flash-UNCENSORED-FP8-BF16.gguf, the vision projector (BF16, converted with llama.cpp convert_hf_to_gguf.py --mmproj), used by flash3090 for image input.

Use with flash3090

git clone https://github.com/ily-lab/flash3090 && cd flash3090   # build steps in its README
hf download ily-lab/GLM-5.3-Flash-UNCENSORED-R2-40-GGUF --local-dir models/experts
hf download dealignai/GLM-5.3-Flash-UNCENSORED-FP8 --local-dir models/GLM-5.3-Flash-FP8
FLASH3090_GPU1=1 tools/flash3090-serve.sh     # 2× RTX 3090; omit FLASH3090_GPU1 for one GPU

The expert file is also a valid ik_llama.cpp GGUF of the full model (it contains all tensors).

License

GLM-5.3-Flash-UNCENSORED — R2-40 IQ5_KS expert GGUF (for flash3090)

Base: zai-org/GLM-5.3-Flash (MIT), uncensored derivative by dealignai (MIT). This is an uncensored model: safety refusals were removed by the derivative's author. Use responsibly.

Quantization: ik_llama.cpp llama-quantize, routed experts IQ5_KS (last MoE layer IQ6_K), shared experts / dense FFN Q8_0, attention and indexer BF16; ~207 GB in 13 shards; importance matrix by AesSedai. Full recipe:

Quality: mean KL divergence vs BF16 of 0.026–0.036 on held-out text, on par with ik_llama.cpp on the same file.

Also included: mmproj-GLM-5.3-Flash-UNCENSORED-FP8-BF16.gguf, the vision projector (BF16, converted with llama.cpp convert_hf_to_gguf.py --mmproj), used by flash3090 for image input.

Use with flash3090

git clone https://github.com/ily-lab/flash3090 && cd flash3090 # build steps in its README hf download ily-lab/GLM-5.3-Flash-UNCENSORED-R2-40-GGUF --local-dir models/experts hf download dealignai/GLM-5.3-Flash-UNCENSORED-FP8 --local-dir models/GLM-5.3-Flash-FP8 FLASH3090_GPU1=1 tools/flash3090-serve.sh # 2× RTX 3090; omit FLASH3090_GPU1 for one GPU

The expert file is also a valid ik_llama.cpp GGUF of the full model (it contains all tensors).