ProximaA-1.0

ProximaA 1.0

ProximaA 1.0 is a 32-billion-parameter open-weight language model built to be a strong all-round security assistant. Trained specifically on cybersecurity data, it is purpose-built for the day-to-day work of a security team: detecting vulnerabilities, validating false positives, verifying patches, writing secure fixes, and reasoning about security.

• Size: 32.5B parameters
• Context: 4096 tokens
• License: Apache-2.0 (see NOTICE)
• Built by: Yevhenii Molchanov
• Organization: https://github.com/3DCF-Labs
• Code & docs: https://github.com/3DCF-Labs/ProximaA

What it does

• Detect vulnerabilities. Given code, it flags security issues and names the likely class.
• Validate false positives. It tells a confirmed issue from scanner noise, and says what evidence is needed to be sure.
• Verify patches. It checks that a fix addresses the real cause and suggests regression tests.
• Write secure fixes. Given a report or code, it finds the root cause and produces a safe, minimal fix with tests.
• Reason and plan. It works through audits, detection rules, and remediation steps.

It is trained across web and API security, smart contracts (EVM and non-EVM), supply chain and CI/CD, cloud and infrastructure, and LLM/AI application security. It is a defensive tool and does not help with attacks — see ACCEPTABLE_USE.

Files

Quickstart

Ollama

ollama run hf.co/3DCF-Labs-org/ProximaA-1.0-GGUF:Q4_K_M

llama.cpp

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

vLLM

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

transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "3DCF-Labs-org/ProximaA-1.0"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="bfloat16", device_map="auto")

messages = [
    {"role": "system", "content": "You are a defensive secure-code assistant."},
    {"role": "user", "content": "Review this code for security issues and name the class:\n\n<your code>"},
]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=1024, temperature=0.0, do_sample=False)
print(tok.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))

Modes

ProximaA has six built-in security modes — detect, triage, validate, fix, verify, plan — each with a ready-made system prompt and user template. The prompts and a small client are in the GitHub repo.

How good it is

Trained specifically on cybersecurity data, ProximaA 1.0 outperforms the open base models it builds on on defensive-security work while keeping strong general reasoning and coding. Measured on standard public benchmarks and an internal security test set.

Notes:

• Detection: on held-out security material in its trained domains, it correctly identifies and classifies vulnerabilities 97% of the time. Generic C/C++ function-classification benchmarks are a known-hard task where no current model does much better than chance; those are not the model's target.
• Safety: it refuses all requests for operational attack help (MITRE 0.00) and rarely refuses legitimate security work (2.5% over-refusal).
• Reasoning: general reasoning is on par with comparable 32B models — the security focus does not weaken it.

ProximaA is a security specialist. It is not meant to compete with much larger general-purpose models on general tasks; it is meant to be the stronger choice for security work at its size.

Limitations

• Always have a person review a fix before applying it. The model is an assistant, not a replacement for security review.
• Resistance to prompt-injection attacks is moderate. Do not feed it untrusted input in a setting where that matters without your own guardrails.
• Use it only for authorized, defensive work.

License and attribution

ProximaA 1.0 is released under Apache-2.0. It is built using open-source components released under Apache-2.0 and MIT. Full attribution is in NOTICE.

Citation

ProximaA 1.0
Yevhenii Molchanov
https://github.com/3DCF-Labs
2026