{ "source_model": "respanai/span-02", "source_revision": "63cbdd774391f1b7d9bc3829179efe996d0191ce", "base_model": "google/gemma-4-26B-A4B-it", "conversion": "Native NVFP4 E2M1 experts and row-scaled FP8 E4M3 weights dequantized to BF16; no retraining.", "unchanged": "Original BF16 vision encoder, vision projection, normalization weights and router scales.", "format": "Sharded standard safetensors", "dtype": "bfloat16", "shards": 32, "tensors": 1013, "tensor_bytes": 51611872412, "config_changes": { "decision_config": { "prompt": "span" } }, "synthetic_smoke_results": { "text_correct": 42, "text_total": 42, "image_correct": 33, "image_total": 33, "requests": 29 }, "limitations": [ "Native activation quantization is not reproduced.", "Native accuracy/benchmark parity has not been established.", "SystemOne requires a vLLM build containing SpanStrategy.", "Native speculative decoding and GPU-specific deterministic thought cost model are not reproduced." ], "systemone_features": [ "choice", "noul", "score", "number", "automatic_thinking", "forced_thinking", "thought_blending", "reasoning_traces", "token_and_time_budgets" ] }