brahmastra-0.3-GGUF

BRAHMASTRA v0.3 — GGUF (Q4_K_M)

Quantized GGUF weights of Krishnapadala55/brahmastra-0.3 for fast inference on consumer GPUs via Ollama or llama.cpp.

Files

*Measured on NVIDIA RTX PRO 5000 Blackwell.

Quick Start with Ollama

# Download (only the Q4_K_M is needed for production)
huggingface-cli download Krishnapadala55/brahmastra-0.3-GGUF brahmastra-v3.Q4_K_M.gguf --local-dir .

# Create Ollama model
cat > Modelfile <<'EOF'
FROM ./brahmastra-v3.Q4_K_M.gguf
PARAMETER temperature 0.6
PARAMETER top_p 0.95
PARAMETER num_ctx 4096
PARAMETER stop "<|im_end|>"
PARAMETER stop "<|endoftext|>"
SYSTEM """You are BRAHMASTRA, an elite AI-powered DAST security scanner. Use <think>...</think> to reason before each response. Be precise — only confirm vulnerabilities with clear evidence."""
EOF

ollama create brahmastra:0.3 -f Modelfile
ollama run brahmastra:0.3

Quick Start with llama.cpp

./llama-cli \
    -m brahmastra-v3.Q4_K_M.gguf \
    --temp 0.6 \
    --top-p 0.95 \
    -c 4096 \
    -p "Analyze this SQL injection: id=1 OR 1=1"

Use as a DAST AI Backend

The companion BRAHMASTRA scanner uses this model via Ollama on localhost:11434. See main repo: https://github.com/psmkreddy/brahmastra

Benchmarks

See full results on the parent repo: Krishnapadala55/brahmastra-0.3

Headline: +23.9pp accuracy on DAST FP judgment vs v0.2 (p<0.001).

License

Apache 2.0 — same as base.