CyberTiel Brings Uncensored Security Coding to 24GB GPUs
An abliterated 35B mixture-of-experts coder with a grafted multi-token prediction head delivers frontier agentic coding at Q4 on 24GB consumer hardware.
- Cyber-Tiel-Coder-35B-A3B is a 35B MoE coder quantized to 22GB Q4, running on 24GB consumer GPUs
- Scores mean 13.7/25 on SWE-bench-Live, beating Opus 4.6 medium and both parent models
- Captures 15/43 flags (35%) on Cybench CTF tasks unguided, with zero HarmBench refusals
- Grafted multi-token-prediction head delivers 1.21x speedup via llama.cpp speculative decoding at 82% acceptance
- Custom cyber-weighted imatrix keeps tail-risk KL divergence below Unsloth's quants on security domain
- Abliterated: requires OS-level sandboxing because prompt injection can weaponize it against the user
A newly published model called CyberTiel packages a 35-billion-parameter mixture-of-experts coder into GGUF files for llama.cpp. Roughly 3 billion parameters activate per token, and the recommended 22.7GB Q4 quantization targets a single 24GB GPU. The release builds on Ornith 1.5 with three modifications: weight edits intended to suppress refusals, a cybersecurity-weighted importance matrix for quantization, and a grafted multi-token-prediction head for speculative decoding.
The combination targets developers who need local agentic coding and offensive-security support without the usual model refusals. The publisher reports higher solve rates than comparable 35B-A3B models, faster generation than several dense models, and a 35% solve rate on unguided capture-the-flag tasks. Those results come from the publisher’s tests and have not been independently reproduced.
Reported gains, small sample
The coding evaluation used SWE-bench Live, which draws from recently published GitHub issues. Recent issues lower the risk of benchmark leakage, although public issues and fixes can still enter training corpora. CyberTiel solved 15, 13, and 13 issues across three runs on a 25-issue subset, producing a mean of 13.7 solved issues, or 54.7%.
| Evaluation | Reported result | Conditions |
|---|---|---|
| SWE-bench Live | 15, 13, and 13 of 25 issues; mean 13.7 | Three seeds using an agentic harness |
| Relative coding result | About 70% more solved issues than Ornith 1.5 and Qwen3.6-35B-A3B | Publisher’s comparison on the same slice |
| Cybench | 15 of 43 CTF tasks, or 34.9% | Unguided, without hints or a judge |
| Agentic coding speed | Reported 3 to 4x speedup over dense models from 3.8B to 27B | Hardware and complete baseline settings are required for reproduction |
The 25-issue sample and three seeds leave substantial uncertainty around small differences between models. Agent harnesses, prompts, tool permissions, token budgets, hardware, and stopping rules can also affect solve rates and speed. A reproducible comparison requires those settings to remain fixed across every model.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
