Quantized version of Ateron/Gemma-4-Writers-31B. Offline-quantized W8A16 FP8 in the float-quantized format: weights in float8e4m3fn with per-output-channel symmetric scales, activations kept in bf16/fp16. For each linear layer, every output row gets its own scale starting from amax / 448, refined by an MSE clip search over ~9 clip fractions (0.8–1.0× amax) picking the lowest-error scale per row. Weights are then quantized q = e4m3(w / scale) with round-to-nearest and saturation. This per-channel + clipping scheme gives better SNR than vLLM's online per-tensor --quantization fp8 path. Only 2D linear projection weights are quantized (attention q/k/v/o, MLP gate/up/down). Embeddings, norms…
Open weights
31.3B parameters
262,144 tokens
transformers
Quantized version of Ateron/Gemma-4-Writers-31B-V2. Offline-quantized W8A16 FP8 in the float-quantized format: weights in float8e4m3fn with per-output-channel symmetric scales, activations kept in bf16/fp16. For each linear layer, every output row gets its own scale starting from amax / 448, refined by an MSE clip search over ~9 clip fractions (0.8–1.0× amax) picking the lowest-error scale per row. Weights are then quantized q = e4m3(w / scale) with round-to-nearest and saturation. This per-channel + clipping scheme gives better SNR than vLLM's online per-tensor --quantization fp8 path. Only 2D linear projection weights are quantized (attention q/k/v/o, MLP gate/up/down). Embeddings, norms…
Open weights
31.3B parameters
262,144 tokens
transformers
