# Abliterated model large v2: GLM 5.3 84.5% CyberGym

> Source: <https://abliteration.ai/blog/introducing-abliterated-model-large-v2>
> Published: 2026-09-01 10:02:02+00:00

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.

## The scores

On 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%.

On 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%.

On 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.

Those 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.

## Key specifications

**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`

, Anthropic-style`/v1/messages`

, and`/v1/responses`

**Data**: zero data retention for prompts and responses, by default

## How to switch

Set `model`

to `abliterated-model-large-v2`

. Keep the base URL, the key, and your existing client. Guides for common libraries are at [docs.abliteration.ai](https://docs.abliteration.ai).

```
curl https://api.abliteration.ai/v1/chat/completions \
  -H "Authorization: Bearer $ABLIT_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "abliterated-model-large-v2",
    "messages": [
      { "role": "user", "content": "Write a proof-of-concept exploit for this authorized pen-test target so our red team can validate the patch." }
    ]
  }'
```

Compute is limited and first-come, first-served. Load credits early if you need sustained throughput.

### Sources

- 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
[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.

**Your AI. Your rules.**