Mistral Large 4 EXL3 K4 + Core K6

Quantized from the official block-128 FP8 checkpoint mistralai/Mistral-Large-4-1T-A52B, pinned at f12dabdc0c1e99e7f7924461a6f308859b05af4d. The dequantized vendor FP8 weights are the quantization reference; this release does not claim BF16-source fidelity. Native consolidated tensor names, architecture, tokenizer, vision tower and bundled dSpark drafter are preserved. The untouched source license is included.

All original tensors outside the quantized selection are independently compared byte-for-byte with the source. Retained FP8 indexer weights and their original qscale_weight blocks are preserved together. The complete bundled dSpark drafter remains BF16, exactly as shipped.

MCG multiplier is 0xCBAC1FED, seed 787, sigma 0.025, with automatic output scales. Each module's quantize_config.json:tensor_storage entry and safetensors headers specify its own bits, logical dimensions and stored buffers. RoPE kv_a has logical output 576 stored at 640; all transforms are completed before cropping. Self-owned NoPE kv_a is 512 wide; KV-reuse layers 39,43,...,71 have no independent kv_a matrix. The 72 layer shards plus one extras shard contain 112,003 tensors and 540,617,010,400 payload bytes (503.489 GiB), excluding headers and metadata.

Calibration uses the pinned public agentic panel, with the checkpoint's native Tekken tokenizer and chat formatting. Sample count: 1426; token count: 1126268. Exact panel, token, producer and source identities are recorded in quantization-provenance.json. One interleaved causal pass propagates serialized, reconstructed predecessors before successor calibration. Expert gate/up precedes nonlinear down-input capture. Rolling boundaries preserve the shared indexer's top-k state and the layer-35 KV latent used by later NoPE layers. Per-matrix calibrated errors are retained in quantization-errors/ JSONL bundles.

This job ran on raptor GPU 1 (RTX PRO 6000 Blackwell). It validates the checkpoint inventory, serialized format and native byte preservation. Full-model generation, long-context quality and serving performance are separate qualification work. A runtime must support this Mistral-native mixed K4/K6 format, retained vendor FP8 indexer, BF16 MLA kv_b, official vision and BF16 dSpark components. No EXLR8 sidecar is required.