Ateron/Gemma-4-Dark-Thoughts-V2-31B
LoRA-style vs ReadyArt/Dark-Scarlett-v2.0-31B. Looks like a merged LoRA (or a fine-tune with frozen norms): only the weight matrices of down_proj, gate_proj, k_proj, o_proj, q_proj, up_proj, v_proj changed, in 60 of 60 layers; norms, biases and embeddings are untouched.
Tensors changed: Embeddings 0/4, Norms 0/523, other 0/62, MLP 180/261, Attention 230/338
8 derivatives checked
- mradermacher/Gemma-4-Dark-Thoughts-V2-31B-i1-GGUF: Quantization (fewer layers)
- Nimbz/Split-31B: LoRA-style
- mradermacher/Gemma-4-Dark-Thoughts-V2-31B-GGUF: Quantization (fewer layers)
- Blazed-Forge/Split-Untied-31B: LoRA-style
- McG-221/Gemma-4-Dark-Thoughts-V2-31B-mlx-8Bit: Quantization
- jancirnodziewiaty/Gemma-4-Dark-Thoughts-V2-31B-oQ4e-mtp: Quantization
- s1arsky/Gemma-4-Dark-Thoughts-V2-31B-exl3-4.0bpw: Different architecture
- nesymerp1/Gemma-4-Dark-Thoughts-V2-31B-Q4_K_M-Q5_K_S: Quantization (fewer layers)