wrldsuksgo2mars/Step-5-Preview-EXL3-K4-Core-K6

Step-5 Preview EXL3 K4 + Core K6

Quantized from the original BF16 release stepfun-ai/Step-5-Preview-BF16, pinned at 7734b7ffac118d70d8398bfa3eb0fab007eb45b1. The checkpoint retains its native sliding-window and sparse full-attention schedule, FP32 residual/norm numerics, headwise attention gates, three MTP stages, bundled vision tower, tokenizer, image processing assets and the source license (copied verbatim).

Part	Matrices	Format
Target attention q/k/v/o, shared/dense FFNs and lm_head	645	EXL3 MCG K6
Routed expert gate/up/down, layers 3–87	89,760	EXL3 MCG K4
Routed expert gate/up/down, layers 88–90	3,168	EXL3 MCG K6
MTP attention, dense FFNs, eh_proj and three stage-local output heads	27	EXL3 MCG K4
Vision, attention gates, sparse indexers, embeddings, routers and norms	1,513 retained tensors	native BF16/F32

MCG multiplier is 0xCBAC1FED, seed 787, sigma 0.025 and automatic output scales. Each of the 352 fused source experts is exported as a separately named standard EXL3 projection under model.layers.N.moe.experts.E in original expert-index order. Each projection contains trellis int16, suh/svh float16 and one int32 MCG multiplier. Per-module metadata in quantize_config.json:tensor_storage and safetensors headers describe each projection's bit width and layout. Target payload is 315,357,147,520 bytes (293.699 GiB), 96 shards and 375,913 tensors. Headers and metadata add overhead. Vision is preserved byte-for-byte and no vision_bits is declared.

Calibration uses the pinned public wrldsuksgo2mars/agentic-quant-panel-v1 with the checkpoint's native chat/thinking rendering. Exact sample/token counts and rendered-token/input/producer hashes are recorded in provenance. Prompts retain original order, batch one, without padding, packing or truncation. Dense Hessians contain actual observations; expert coverage and recovery evidence are recorded separately.

One interleaved causal pass propagates serialized/reconstructed core and expert predecessors before calibrating each successor layer. Gate/up precedes the native clamped SwiGLU down-input capture. Sliding-window attention uses the native 512-token causal window, with the native sparse indexer on the 23 full-attention layers. All three MTP stages use their native shifted-token alignment and retain their own calibrated K4 output head.

Quantization ran on maxq's single RTX PRO 6000 Blackwell Max-Q. Every retained tensor was independently compared byte-for-byte with the BF16 source; packed layouts and the complete inventory were checked. Evidence is in quantization-errors/, with producer/input hashes in quantization-provenance.json. Full-model generation, long-context quality and serving performance qualification are separate from this job.

A runtime must support Step-5's mixed standard MCG K4/K6 expert and core projections, three MTP stages and retained native components. No EXLR8 sidecar is required.

Native tokenizer

Calibration used the exact bundled ByteLevel tokenizer backend and the native max-effort chat template. Load the tokenizer with PreTrainedTokenizerFast.from_pretrained(...) to preserve that backend. The upstream tokenizer configuration carries a historical Llama class label; family-specific AutoTokenizer conversion in some Transformers versions changes the backend. The original tokenizer assets are preserved byte-for-byte here.