ProximaA 1.0

A 32B open-weight security assistant. Trained specifically on cybersecurity data for the full defensive workflow: vulnerability detection, false-positive validation, patch verification, secure fixes, and security reasoning.

Built by Yevh, 3DCF Labs

• Weights & GGUF: Hugging Face
• License: Apache-2.0 (see NOTICE.md)

---

What it does

• Detect vulnerabilities in code and name the likely class.
• Validate false positives — tell a real issue from scanner noise.
• Verify patches — check a fix addresses the real cause; suggest regression tests.
• Fix — find the root cause and produce a safe, minimal patch.
• Reason and plan — audits, detection rules, remediation steps.

It is trained for defenders. It does not help with attacks. See docs/ACCEPTABLE_USE.md.

Get the model

The weights and quantized GGUF files live on Hugging Face. Pick a runtime:

Ollama (laptop / desktop)

# quantized build, no setup
ollama run hf.co/3DCF-Labs-org/ProximaA-1.0-GGUF:Q4_K_M

Or build locally from the included Modelfile after downloading a GGUF:

ollama create proximaa -f Modelfile
ollama run proximaa

llama.cpp (GGUF)

llama-server -m proximaa-1.0-Q4_K_M.gguf -c 4096

vLLM (GPU, full precision)

vllm serve 3DCF-Labs-org/ProximaA-1.0 --dtype bfloat16 --max-model-len 4096

Will it run on a laptop?

The Q4_K_M GGUF is ~19 GB. It runs well on machines with 32 GB+ of unified/​system memory (Apple Silicon M-series is a good fit). On 16 GB laptops it will be tight to unusable — use a smaller machine tier or a GPU host instead.

Modes

ProximaA ships with six built-in security modes — detect, triage, validate, fix, verify, plan — each a ready-made system prompt plus user template in prompts/modes.json. See docs/MODES.md.

# examples/proximaa_client.py — loads a mode and calls any OpenAI-compatible server
python examples/proximaa_client.py --mode fix --input report.txt
python examples/proximaa_client.py --mode detect --input snippet.c

There is also a plain examples/curl_example.sh.

Repository layout

README.md              this file
Modelfile              Ollama build file (points at a local GGUF)
prompts/               default system prompt + the six mode prompts (modes.json)
docs/                  MODEL_CARD, MODES, USAGE, SAFETY, ACCEPTABLE_USE
examples/              Python mode client, curl example, sample interactions
NOTICE.md              license and attribution
LICENSE-APACHE.txt     Apache-2.0
LICENSE-MIT.txt        MIT (for components that require it — see NOTICE.md)

Documentation

• docs/MODEL_CARD.md — what the model is, what it's good at, benchmark scores.
• docs/MODES.md — the six modes and output formats.
• docs/USAGE.md — how to run it, with examples.
• docs/ACCEPTABLE_USE.md and docs/SAFETY.md — the rules.

Citation

ProximaA 1.0
Yevh, 3DCF Labs
https://github.com/3DCF-Labs
2026