anikifoss/DeepSeek-V3.1-HQ4_K

High quality quantization of DeepSeek-V3.1 without using imatrix.

The architecture has not changed, so token generation speed should be the same as DeepSeek-R1-0528, see benchmarks here.

Run

ik_llama.cpp

See this detailed guide on how to setup ik_llama and how to make custom quants.

./build/bin/llama-server \
    --alias anikifoss/DeepSeek-V3.1-HQ4_K \
    --model /home/gamer/Env/models/anikifoss/DeepSeek-V3.1-HQ4_K/DeepSeek-V3.1-HQ4_K-00001-of-00010.gguf \
    --no-mmap \
    --temp 0.5 --top-k 0 --top-p 1.0 --min-p 0.1 --repeat-penalty 1.0 \
    --ctx-size 82000 \
    -ctk f16 \
    -mla 3 -fa \
    -amb 512 \
    -b 1024 -ub 1024 \
    -fmoe \
    --n-gpu-layers 99 \
    --override-tensor exps=CPU \
    --parallel 1 \
    --threads 32 \
    --threads-batch 64 \
    --host 127.0.0.1 \
    --port 8090

llama.cpp

You can turn on thinking by changing "thinking": false to "thinking": true below.

Currently llama.cpp does not return <think> token in response. If you know how to fix that, please share in the "Community" section!

As a workaround, to inject the token in OpenWebUI, you can use the inject_think_token_filter.txt code included in the repository. You can add filters via Admin Panel -> Functions -> Filter -> + button on the right

./build/bin/llama-server \
    --alias anikifoss/DeepSeek-V3.1-HQ4_K \
    --model /home/gamer/Env/models/anikifoss/DeepSeek-V3.1-HQ4_K/DeepSeek-V3.1-HQ4_K-00001-of-00010.gguf \
    --temp 0.5 --top-k 0 --top-p 1.0 --min-p 0.1 --repeat-penalty 1.0 \
    --ctx-size 64000 \
    -ctk f16 \
    -fa \
    --chat-template-kwargs '{"thinking": false }' \
    -b 1024 -ub 1024 \
    --n-gpu-layers 99 \
    --override-tensor exps=CPU \
    --parallel 1 \
    --threads 32 \
    --threads-batch 64 \
    --jinja \
    --host 127.0.0.1 \
    --port 8090