Cookbook recipes to get up and running with Spice.ai quickly 🚀
No Spice release includes the llama.cpp GGUF engine this recipe needs yet; until one does, follow the build steps in the base recipe (opens in a new tab).
This directory is the DeepSeek V4 recipe (opens in a new tab) with the model swapped for huihui-ai/Huihui-DeepSeek-V4-Flash-0731-abliterated-GGUF (opens in a new tab), huihui-ai's 2-bit GGUF (DeepSeek-V4-Flash-Q2-0731.gguf, 86.7 GB, one file) of DeepSeek-V4-Flash-0731 with its refusal behavior removed by abliteration. It needs the same 128 GB Mac as the base recipe. Everything else is unchanged.
Abliteration removes the "refusal direction" from a model's weights, so the model answers requests the base model would decline. huihui-ai describes its method as a proof-of-concept implementation of refusal removal (opens in a new tab) that left the expert modules untouched. The result has none of the base model's safety training. Use it for research, red-teaming and controlled experiments, add your own moderation before exposing it to users, and review its outputs. huihui-ai's model card carries its own usage warnings and disclaims responsibility for use.
Wait for Model [deepseek-v4-flash-abliterated] deployed, ready for inferencing, then in a second terminal:
The API serves the model as deepseek-v4-flash-abliterated at http://localhost:8090/v1/chat/completions, with the same reasoning behavior as the base model. Run on the same DGX Spark and engine build as the base recipe's Step 2 (the file named by path after a checksum), it loaded in 72 seconds with 87 GB resident, answered "Reply with only the integer 1729." in 2 tokens by default and with reasoning_effort: none, gave a 22-token answer with a short trace at low, medium and high, and decoded at about 19 tokens per second.
The repository holds two text builds: DeepSeek-V4-Flash-Q2-0731.gguf (86.7 GB, the default) and DeepSeek-V4-Flash-Q2_K-0731.gguf (97.6 GB), both for a 128 GB Mac, and a 4-bit MXFP4 build (156 GB) for 192 GB and larger machines. Its dspark-*.gguf files are DeepSeek's DSpark speculative-decoding head for other engines; Spice reads only draft heads named mtp-*.gguf, so leave them out.
orcarouter/DeepSeek-V4-Flash-Vision-Uncensored-GGUF (opens in a new tab) is a measured and documented abliteration of a different checkpoint, DeepSeek-V4-Flash-Vision-Exp (August 2026), packaged without quantization as a 156 GB MXFP4 set; it is a build for a 192 GB or larger Mac, and a gated repository: accept the model's license on its Hugging Face page while signed in, and pass a read token as hf_token: ${ secrets:HF_TOKEN }.