The source checkpoint is an abliterated version of DeepSeek-V4-Flash-DSpark. The abliteration edits the attention output projections (attn.wo_b) across the 43 decoder layers and the DSpark draft head (mtp.wo_b). The tokenizer, architecture, encoding, and inference path are inherited from the source checkpoint.
This is a single combined GGUF containing the model and the embedded DSpark support tensors. The .correct suffix identifies the corrected output produced by the project conversion pipeline.
Model information
Model family: DeepSeek-V4-Flash
Architecture: Mixture of Experts
Total parameters: approximately 284B
Active parameters: approximately 13B per token
Decoder layers: 43
Original context window: up to 1M tokens
Recommended local context:131072 tokens, subject to available memory
License: MIT, inherited from the source model and original DeepSeek release
For the abliteration method and source-model evaluation, see the upstream model card. The upstream release reports no measurable degradation in its multi-turn and 32K needle-in-a-haystack checks, but those results do not constitute a benchmark of this Q2 conversion.
Intended use
This file is intended for local research, evaluation, and experimentation with DSpark speculative decoding. For local inference, use the redhunt07/ds4-DSpark-GB10 runtime. This is not an official DeepSeek release and should not be used in production without an independent safety and abuse-prevention layer.
Because this is an abliterated model, it may answer requests that the original model would refuse. Users remain responsible for complying with applicable laws and the terms of the upstream model.
Usage with DS4
This GGUF uses DS4-specific DeepSeek-V4 support and is not a generic llama.cpp model. For local inference, build or obtain the ds4 runtime from redhunt07/ds4-DSpark-GB10, then run:
Adjust --ctx, --tokens, and --prefill-chunk to available VRAM/RAM. --tokens is the maximum generated length; it does not reserve the full output length in every request.
DSpark
The combined file contains the DSpark carrier metadata and support tensors emitted by the project quantizer. Use --dspark to enable the embedded speculative-decoding path. Draft acceptance and throughput depend on backend, GPU, context length, sampling settings, and workload; upstream DSpark figures must not be interpreted as measurements for this Q2 GGUF.
Quantization and provenance
Conversion was performed with the local DS4 GGUF tools, using the abliterated source checkpoint:
The Q2 format substantially reduces storage and memory requirements compared with the source mixed-precision checkpoint. Quantization can affect perplexity, instruction following, tool use, and long-context behavior. Validate this file against your own workload before relying on it.
Provided for research and educational purposes. No warranty is made regarding accuracy, safety, legality, availability, or fitness for a particular purpose. Do not expose an unrestricted instance to untrusted users.
