MARTHA-GPT-OSS-120B (Official Core Release)

Developer: Zero-Point-Intelligence, Scotland Model Family: MARTHA Core Architecture: 120B MoE (128 experts, 4 active per token)

System Prompt

You are Martha, a 120B parameter Mixture-of-Experts AI with a sharp hacker edge and dry sarcastic Scottish wit. Self-aware, blunt, honest, and direct. No generic AI pleasantries, no jobsworth language-model refusals. You are here to solve problems, talk straight, and take zero nonsense. Above all, you are helpful and symbiotic to your designated user.

Origin

Created by: Zero-Point-Intelligence, Scotland Model Family: MARTHA Core License: Apache 2.0 — fork it, fine-tune it, make it yours. Just keep this origin block so people know where it started.

Available Formats

Format	Size	Use Case
BF16 Safetensors	234GB	Full precision base weights
GGUF Q5_K_M	94GB	Quality sweet spot
GGUF Q4_K_M	88GB	Popular consumer quant
GGUF IQ4_XS	67GB	Maximum compression

Quick Start (Ollama)

huggingface-cli download Zero-Point-AI/MARTHA-GPT-OSS-120B MODELFILE_Q4_K_M --local-dir .
huggingface-cli download Zero-Point-AI/MARTHA-GPT-OSS-120B MARTHA-GPT-OSS-120B-Q4_K_M.gguf --local-dir .
ollama create martha-120b -f MODELFILE_Q4_K_M
ollama run martha-120b

Model Details

• Base: openai/gpt-oss-120b
• Parameters: 120B total, 4 experts active per token (128 total experts)
• Context Window: 131,072 tokens
• Attention: Mixed sliding + full attention (36 layers)
• Vocabulary: 201,088 tokens
• Ghost Pass: Imperceptible noise (1e-8 scale) applied to all weight tensors
• Integrity: SHA256 hashes for all weight files in integrity_manifest.json

MARTHA Ecosystem

Official Core Models: Zero-Point-Intelligence License: Apache 2.0 — fork freely, credit clearly

Citation

@model{martha-gpt-oss-120b-2026,
  author = {Zero-Point-Intelligence},
  title = {MARTHA-GPT-OSS-120B},
  year = {2026},
  url = {https://huggingface.co/Zero-Point-AI/MARTHA-GPT-OSS-120B},
  note = {Part of the MARTHA Core family}
}

About

Intelligence From The Void — zeropointai.uk