Blackfrost-AI/RED-SNOW-5.3-FLASH-BF16
LoRA-style vs zai-org/GLM-5.3-Flash-BF16. Looks like a merged LoRA (or a fine-tune with frozen norms): only the weight matrices of b_proj, down_proj, f_a_proj, f_b_proj, g_a_proj, g_b_proj, gate_proj, k_proj, kv_a_proj_with_mqa, o_proj, q_a_proj, q_b_proj, q_proj, up_proj, v_proj changed, in 45 of 46 layers; norms, biases and embeddings are untouched.
Tensors changed: Attention 350/827, MLP 605/1888, lm_head 0/1, Embeddings 0/3, Norms 0/277, other 0/4
5 derivatives checked
- Blackfrost-AI/RED-SNOW-5.3-FLASH-FP8: Quantization
- freakyskittle/RED-SNOW-5.3-FLASH-GGUF: Quantization
- vcruz305/RED-SNOW-5.3-FLASH-EXL3-SAGE-4.91bpw: Different architecture
- vcruz305/RED-SNOW-5.3-FLASH-EXL3-SAGE-2.49bpw: Different architecture
- Blackfrost-AI/RED-SNOW-5.3-FLASH-NVFP4: Quantization