{ "generated_from": "kv-bytes-per-token.json (fixtures computed from pinned HF config.json revisions read 2026-11-03) and GGUF LFS sizes from the same run", "formula": "dense: 2 * n_layers * n_kv_heads * head_dim * bytes_per_element * tokens; hybrid: 2 * full_layers * global_kv_heads * global_head_dim * bpe * tokens + 2 * sliding_layers * n_kv_heads * head_dim * bpe * min(tokens, window)", "bytes_per_element": {"f16": 2.0, "q8_0": 1.0625}, "models": { "qwen3-8b": { "repo": "Qwen/Qwen3-8B", "revision": "b968826d9c46dd6066d109eabc6255188de91218", "num_hidden_layers": 36, "num_key_value_heads": 8, "head_dim": 128, "max_position_embeddings": 40960, "gguf_gib": {"Q4_K_M": 4.68, "Q8_0": 8.11, "BF16": 15.26} }, "qwen3-14b": { "repo": "Qwen/Qwen3-14B", "num_hidden_layers": 40, "num_key_value_heads": 8, "head_dim": 128, "max_position_embeddings": 40960, "gguf_gib": {"Q4_K_M": 8.38, "Q8_0": 14.62, "BF16": 27.51} }, "devstral-small-2507": { "repo": "mistralai/Devstral-Small-2507", "num_hidden_layers": 40, "num_key_value_heads": 8, "head_dim": 128, "max_position_embeddings": 131072, "gguf_gib": {"Q4_K_M": 13.35, "Q8_0": 23.33, "BF16": 43.92} }, "gemma-4-31b-it": { "repo": "google/gemma-4-31B-it", "num_hidden_layers": 60, "num_key_value_heads": 16, "head_dim": 256, "max_position_embeddings": 262144, "gguf_gib": {"Q4_K_M": 17.07, "Q8_0": 30.39}, "hybrid": { "full_layers": 10, "sliding_layers": 50, "window": 1024, "global_head_dim": 512, "global_kv_heads": 4 } } } }