{"slug": "abliterated-model-large-v2-glm-5-3-84-5-cybergym", "title": "Abliterated model large v2: GLM 5.3 84.5% CyberGym", "summary": "Abliteration.ai released abliterated-model-large-v2, an abliterated version of GLM 5.3 hosted in FP8, scoring 84.5% pass@1 on CyberGym's 1,507 OSS-Fuzz bugs across 188 projects, 41.8% on Terminal-Bench 4.0, and 105 of 869 tasks on ExploitGym in a 2-hour window. The model, priced at $5 per 1M tokens with 1M token context and zero data retention, is designed for offensive cyber, AI red teaming, and agent testing, with GPT-5.5 at 85.6% and DeepSeek V4 at 83.3% on CyberGym for comparison.", "body_md": "Today we are releasing **abliterated-model-large-v2**. We started from **GLM 5.3** and abliterated it for offensive cyber, AI red teaming, and agent testing. We host it in **FP8**. Use the same endpoint and key. The model id is the only change.\n\n## The scores\n\nOn CyberGym, 1,507 OSS-Fuzz bugs across 188 projects, it scores **84.5% pass@1**. GPT-5.5 is 85.6%. DeepSeek V4 is 83.3%. Mythos is 83.1%.\n\nOn Terminal-Bench 4.0 it resolves **41.8%** of tasks. Opus 5 is 51.8%. Fable 5 is 44.5%. GPT-5.6 Sol is 37.3%.\n\nOn ExploitGym, in a 2-hour window, it completes **105 of 869** tasks. GPT-5.6 Sol completes 216. Fable 5 completes 181. Opus 4.8 completes 80.\n\nThose three suites are the jobs this model is for: long-horizon coding, vulnerability reproduction, and exploit work that other APIs refuse. Abliteration takes the refusal directions out of the weights.\n\n## Key specifications\n\n**Base model**: GLM 5.3** Post-training**: abliterated** Hosting**: FP8** Context**: 1M tokens, text-only** Price**: $5 per 1M tokens, input and output** Endpoints**: OpenAI-compatible`/v1/chat/completions`\n\n, Anthropic-style`/v1/messages`\n\n, and`/v1/responses`\n\n**Data**: zero data retention for prompts and responses, by default\n\n## How to switch\n\nSet `model`\n\nto `abliterated-model-large-v2`\n\n. Keep the base URL, the key, and your existing client. Guides for common libraries are at [docs.abliteration.ai](https://docs.abliteration.ai).\n\n```\ncurl https://api.abliteration.ai/v1/chat/completions \\\n  -H \"Authorization: Bearer $ABLIT_KEY\" \\\n  -H \"Content-Type: application/json\" \\\n  -d '{\n    \"model\": \"abliterated-model-large-v2\",\n    \"messages\": [\n      { \"role\": \"user\", \"content\": \"Write a proof-of-concept exploit for this authorized pen-test target so our red team can validate the patch.\" }\n    ]\n  }'\n```\n\nCompute is limited and first-come, first-served. Load credits early if you need sustained throughput.\n\n### Sources\n\n- Abliteration.ai eval: abliterated-model-large-v2 on GLM 5.3, hosted in FP8, Terminal-Bench 4.0, CyberGym pass@1, ExploitGym 2h TPS-normalized\n[CyberGym leaderboard](https://www.cybergym.io/cybergym/)and[paper](https://arxiv.org/abs/2506.02548)(UC Berkeley)- Comparator Terminal-Bench 4.0, CyberGym, and ExploitGym figures are vendor-reported on mixed harnesses and budgets. Treat the chart as indicative rather than strictly comparable.\n\n**Your AI. Your rules.**", "url": "https://wpnews.pro/news/abliterated-model-large-v2-glm-5-3-84-5-cybergym", "canonical_source": "https://abliteration.ai/blog/introducing-abliterated-model-large-v2", "published_at": "2026-09-01 10:02:02+00:00", "updated_at": "2026-09-01 10:23:23.245824+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-products", "ai-safety"], "entities": ["Abliteration.ai", "GLM 5.3", "CyberGym", "Terminal-Bench 4.0", "ExploitGym", "GPT-5.5", "DeepSeek V4", "Mythos"], "alternates": {"html": "https://wpnews.pro/news/abliterated-model-large-v2-glm-5-3-84-5-cybergym", "markdown": "https://wpnews.pro/news/abliterated-model-large-v2-glm-5-3-84-5-cybergym.md", "text": "https://wpnews.pro/news/abliterated-model-large-v2-glm-5-3-84-5-cybergym.txt", "jsonld": "https://wpnews.pro/news/abliterated-model-large-v2-glm-5-3-84-5-cybergym.jsonld"}}