● ARCH: DEEPSEEK4 (256×8.4B MOE) ▲ HF OAUTH: CHECKING... READY
MODEL CARD ↗ ANTIREZ/DS4 ↗ LLAMA.CPP-DS4-CUDA ↗
HUIHUI-AI // ABLITERATED DWARFSTAR (DS4) & LLAMA.CPP QUANTIZATION

Huihui-DeepSeek-V4-Flash-abliterated-ds4-GGUF

Uncensored abliterated GGUF suite of deepseek-ai/DeepSeek-V4-Flash engineered for DwarfStar (ds4) and CUDA/Metal workstations. Stream cloud comparisons via 1-click Hugging Face OAuth, connect directly to your own ds4-server, or inspect the live GGUF v3 tensor headers via HTTP Range.

▲ UPSTREAM USAGE WARNING
Safety filtering has been significantly reduced via abliteration. Intended for controlled research, benchmarking, and local engineering workflows.
MOE TOPOLOGY
256experts (top-6 + 1 shared)
TRANSFORMER BLOCKS
43layers + 1 MTP
NATIVE CONTEXT
1,048,576tokens (YaRN ×16)
SPARSE ATTENTION
MLA + DSAtop-512 indexer
Q2 FOOTPRINT
80.76GiB (IQ2_XXS/Q2_K)
RTX PRO 6000 (96GB)
>35tok/s (ds4 CUDA)
Streaming Session Output SSE REAL-TIME TELEMETRY
Live HTTP Range (206) GGUF v3 Binary Header & Tensor Parser IDLE
GGUF Architecture Metadata KV Pairs
Metadata Key Parsed Binary Value
Embedded Jinja Chat Template (tokenizer.chat_template)

              
Tensor Graph Layout Explorer
# Tensor Name Quant Dimensions Data Offset
Memory Budget & Generated Launch Commands ● RESIDENT FIT
ESTIMATED RESIDENT FOOTPRINT
88.96GiB / 96 GiB
1. Download Selected GGUF Checkpoints

              
2. Build & Serve with DwarfStar (antirez/ds4 or huihui-support/ds4 TP/UDS)

              
3. Alternative: Fringe210/llama.cpp-deepseek-v4-flash-cuda