Tkaa/Merkaba01-Flash1017-Tier2

Merkaba01 — DeepSeek-V4.1-Flash-1017 Tier-2 Runtime Core Package (v0.1 Alpha)

This repository contains the pre-packaged Merkaba01 Tier-2 runtime weights derived from DeepSeek-V4.1-Flash-1017 (NVFP4) (nvidia/DeepSeek-V4.1-Flash-1017-NVFP4, 40 layers, d_model = 5120, Multi-head Latent Attention MLA, Mixture-of-Experts MoE, Engram N-gram lookup, and Multi-Token Prediction MTP), including the pre-built merkaba_backbone_q4.bin and merkaba_mtp0_q4.bin hardware-acceleration sidecars.

Note: These files use Merkaba01's custom Tier-2 expert layout and binary sidecar formats (.bin) and are designed exclusively for the Merkaba01 Pure-Rust Cognitive Runtime Engine (merkaba-gemma chat flash / flash1017-run).

🔒 Hardware Lock (MERKABA_GB10_LOCK_V1 — GB10 / DGX Spark Exclusive): Layer normalization and routing tensors in this package are cryptographically bound to NVIDIA Grace Blackwell (GB10 / DGX Spark, Linux aarch64, Compute Capability 12.1 / sm_121) via addon_gx10_rust. Execution on Windows/RTX, Apple Silicon (Mac), Intel, or AMD is disabled in this release. (Svenska: Denna modell är hårdvarulåst exklusivt till NVIDIA Grace Blackwell GB10 / DGX Spark. Körning på Windows/RTX, Mac eller AMD är spärrad och kräver officiellt hårdvarupartnerskap/sponsring).

---

📦 Included Files in This Repository (~105.1 GiB)

• model-00001-of-00048.safetensors .. model-00044-of-00048.safetensors (44 shards, ~92.7 GiB) — Tier-2 core MoE experts and MLA weights (NVFP4 / BF16).
• merkaba_backbone_q4.bin (12.2 GiB) — Pre-built zero-copy mmap Q4_G32 MLA & shared-backbone sidecar.
• merkaba_mtp0_q4.bin (82 MiB) — Pre-built Q4_G32 Multi-Token Prediction (MTP Layer 40) speculative draft sidecar.
• merkaba_dspark_heads.safetensors (126 MiB) — Calibrated DSpark / SilentVectorLoop (JEPA) and CoarseToFineConceptHead (C2F) heads.
• config.json, hf_quant_config.json, model.safetensors.index.json, tokenizer.json, tokenizer_config.json — Model configuration, tensor shard index, and tokenizer vocabulary.

---

⚠️ Important: Where Are Shards 00045–00048? (Engram N-gram Tables)

English

1. Shards 00045 & 00046 are not needed: Because the Tier-2 core expert architecture compacts the MoE layers into 44 shards (00001–00044), shards 00045 and 00046 do not exist and are not referenced by model.safetensors.index.json.
2. Shards model-00047-of-00048.safetensors (48.78 GiB) and model-00048-of-00048.safetensors (48.79 GiB) are fetched from nvidia/DeepSeek-V4.1-Flash-1017-NVFP4:
   • These two shards contain the 100% unmodified Engram N-gram associative embedding tables (used in Layers 1, 14, and 27).
   • Because they are byte-for-byte identical to NVIDIA's public repository, we do not duplicate ~97.6 GiB of identical data here.
   • Automatic Download (download_model.sh): If you use ./download_model.sh from Merkaba01-Flash1017-Tier2-Runtime-v0.1.zip, the script automatically downloads shards 00001–00044 + .bin sidecars from Tkaa/Merkaba01-Flash1017-Tier2 AND downloads model-00047-of-00048.safetensors + model-00048-of-00048.safetensors directly from nvidia/DeepSeek-V4.1-Flash-1017-NVFP4.
   • Manual Download: If you download the files manually from Hugging Face, place all files from this repository into your model/ folder, and then download shards 00047 and 00048 from nvidia/DeepSeek-V4.1-Flash-1017-NVFP4 into the same model/ folder:

     curl -L -C - -o model/model-00047-of-00048.safetensors \
       https://huggingface.co/nvidia/DeepSeek-V4.1-Flash-1017-NVFP4/resolve/main/model-00047-of-00048.safetensors
     
     curl -L -C - -o model/model-00048-of-00048.safetensors \
       https://huggingface.co/nvidia/DeepSeek-V4.1-Flash-1017-NVFP4/resolve/main/model-00048-of-00048.safetensors
     

     (Note: If you want to run on a system with limited disk space without the 97.6 GiB Engram shards, pass --no-engram to merkaba-gemma).

---

🇸🇪 Svenska — Viktig information vid manuell nedladdning (00047 & 00048)

Denna modell bygger på DeepSeek-V4.1-Flash-1017 (NVFP4) anpassad för Merkaba01 Tier-2.

1. Varför model-00045 och model-00046 inte finns: Tier-2-arkitekturen rymmer alla aktiva MoE-experter och MLA-lager i shard 00001 till 00044. Shard 00045 och 00046 används därför inte.
2. Varför model-00047-of-00048.safetensors (48.78 GiB) och model-00048-of-00048.safetensors (48.79 GiB) hämtas från NVIDIA:
   • Dessa två filer innehåller DeepSeek Flash Engram N-gram-tabellerna (Lager 1, 14 och 27) och är 100 % oförändrade jämfört med originalet.
   • Om du använder ./download_model.sh i Merkaba01-Flash1017-Tier2-Runtime-v0.1.zip hämtas shard 00001–00044 automatiskt härifrån (Tkaa/Merkaba01-Flash1017-Tier2) och shard 00047–00048 hämtas automatiskt direkt från det öppna repot nvidia/DeepSeek-V4.1-Flash-1017-NVFP4.
   • Vid manuell nedladdning: Om du laddar ner filerna manuellt, lägg alla filer från detta repo i din model/-mapp och ladda sedan ner model-00047-of-00048.safetensors samt model-00048-of-00048.safetensors från nvidia/DeepSeek-V4.1-Flash-1017-NVFP4/tree/main till samma model/-mapp (eller starta motorn med flaggan --no-engram om du vill köra utan N-gram-tabellerna).