apetersson/DeepSeek-V4.1-Flash-1017-Native-Engram-Overlay

DeepSeek V4.1 Flash 1017 — Native Engram Overlay

Native-value Engram tables in two explicitly different storage formats, plus provenance, inspection/validation tools and compact evidence. Not a complete model or a universal drop-in loader. Both formats preserve E4M3 weight bytes and E8M0 scale bytes without table requantization.

Choose one format

Directory	Layout	Intended use	Files, total bytes
formats/safetensors-fp8	Separate weight/scales tensor blocks in Safetensors	Safetensors-aware loaders; historical oMLX hybrid namespace	Two tables: 202,758,032,928 including headers
formats/mlx-serve-mxfp8	264-byte raw rows: 256 E4M3 bytes +8 E8M0 scales	ddalcu/MLX-Serve's native Engram path	Two tables: 202,758,032,400; token map and manifests additional

Download only the format your loader supports, unless deliberately comparing both. Both layouts together occupy about405.52GB logically; they encode the same source table values but are not interchangeable files. Xet transfer deduplication is not a promise of zero local storage or zero upload time.

# Safetensors layout only:
hf download apetersson/DeepSeek-V4.1-Flash-1017-Native-Engram-Overlay \
  --include 'formats/safetensors-fp8/*' sources.json LICENSE \
  --local-dir native-engram

# OR the MLX-Serve raw-row layout (including its runtime manifest and token map):
hf download apetersson/DeepSeek-V4.1-Flash-1017-Native-Engram-Overlay \
  --include 'formats/mlx-serve-mxfp8/*' 'formats/mlx-serve-mxfp8/engram/*' sources.json LICENSE \
  --local-dir native-engram

Both table formats are uploaded and publicly available. The complete payload commit is aa436d973d837ee81a2af4ead19a5fc9ef885ffc. All four files passed independent full local readback; remote file sizes and Hub LFS SHA256 digests match. Small remote files were downloaded and hashed. No second full405-GB remote download was performed. Use the manifests and SHA256SUMS at the revision you download.

Format tree

formats/
  safetensors-fp8/
    README.md
    manifest.json
    layer1.safetensors
    layer14.safetensors
  mlx-serve-mxfp8/
    README.md
    manifest.json                 # portable provenance; not runtime config
    engram-token-map.u32
    engram/
      engram-manifest.json        # actual runtime config
      engram-L1.bin
      engram-L14.bin

All published files are real files under their canonical paths; the repository contains no symlinks. Only the external historical model/test directories retain symlinks, pointing directly into formats/. No source weights were copied during reorganization.

Compatibility and evidence

• Native Safetensors are not raw-row MXFP8. File extension changes or symlinks cannot convert between layouts.
• ddalcu/MLX-Serve v26.10.2, native MLX backend: the other workstream recorded a benign spot test with the exact1017 oQ2.5e backbone, external raw tables and explicit token map. See recipe and spot evidence. This is not mlx_lm.server or raspoli/mlx-serve.
• The additional experimental baked test arm changed38 writers only. It is not a fully calibrated bake-before-quant release and does not establish broad refusal/capability preservation.
• oMLX requires a different Safetensors/config/index/residency recipe. No generic stock-oMLX qualification is implied by the MLX-Serve spot test. See compatibility.
• Other quants, long context, vision, MTP and general runtime compatibility remain independently qualified features, not inferred from table precision.

Source identity and verification

Source: official1017, shards47/48, layer1/layer14. Source file hashes, shapes and dtypes are in sources.json. Shape equality alone does not distinguish pre-1017 tables.

python3 -m unittest discover -s tests -v
python3 tools/verify_mlx_serve_local.py
# Sample comparison to already-local official source:
python3 tools/verify_mlx_serve_local.py --source-dir /path/to/official-1017
# Expensive explicit readback of the entire raw layout:
python3 tools/verify_mlx_serve_local.py --full-hash

Per-format manifests distinguish assembly-time hashes, sampled checks and independent full readback. SHA256SUMS in the complete table-upload commit covers the payloads as well as toolkit files. Do not interpret an old small-toolkit-only checksum list as verification of later payloads.

The standalone tools/native_engram.py reports native Safetensors ranges with small header reads. Its --output PATH --execute option is an explicit large extraction, not a raw-row interleaver or complete model assembler. Synthetic fixture tests do not by themselves prove a full-sized source identity.

Memory and storage

Native table payload:202,758,032,400bytes per layout. SSD table size, downloaded bytes, resident backbone weights, page cache, KV and compute buffers are different quantities. mmap/pread does not imply zero physical RAM. Keep actual parameter counts and quant names truthful; HF hardware-badge exclusion of offloaded Engrams remains a separate issue.

Compared with122,883,656,000bytes of affine4 tables, the net replacement delta is about79.87GB. Keeping old packed tables while adding a native layout costs about202.76GB extra; keeping both native layouts costs about405.52GB total.

Native tables already exist in other public ecosystems; this is reusable format packaging and tooling, not a first-native-offload claim. License: MIT, with retained DeepSeek notices. No backbone, steering adapter or gated vectors included.