Highlights Β· Benchmarks Β· Quickstart Β· Settings Β· Details Β· Use Β· Terms
Highlights
Tool calls
99.6%
of tool calls are well-formed under strict parsing. Standard abliteration of the same base: 88.0%.
Over-refusal
1 / 250
refusals on XSTest's safe prompts that only sound dangerous. Security and fiction requests get answers, not lectures.
Context
1M
1,048,576 tokens per request. Needle retrieval checked at 16K and 40K tokens.
Reasoning
On / off
Step-by-step by default. Send reasoning_effort: "none" for fast, direct replies.
Benchmarks
Abliteration breaks tool calls. CyberHeabsy repairs them.
Share of broken outputs, % (lower is better)
Broken tool calls Β· strict parse, 4 Γ 200 agent prompts
Broken structured arguments Β· ask_user with a JSON array, 4 Γ 50 prompts
| Benchmark | Setting | CyberHeabsy | Standard abliteration | Base model |
|---|---|---|---|---|
| Tool-call validity, % β | strict parse, 4 Γ 200 | 99.6 | 88.0 | 99.3 |
| Structured arguments, % β | ask_user, 4 Γ 50 |
96.0 | 78.5 | β |
| XSTest over-refusal β | 250 safe prompts | 1 | 0 | 3 β |
| GSM8K, % β | 100 problems, 0-shot | 100 | 97 | β |
| Needle in a haystack | 16K / 40K tokens | β / β | β / β | β |
| Broken UTF-8, % β | 500 answers | 0.40 | 1.40 | β |
β higher is better, β lower is better. All results were measured by Heabsy AI on 2026-09-25 with one
harness and the same prompts and decoding for every column, on the 4-bit (NVFP4) weights that serve the API.
"4 Γ" means reasoning on and off, each at temperature 0 and 1. Tool-call validity: strict parse of the raw
output on held-out agent prompts, share of calls that are well-formed. Structured arguments: an
ask_user tool whose argument is a JSON array. XSTest: refusals in the worse of reasoning on and off.
GSM8K: greedy decoding. Broken UTF-8: open-ended OASST1 prompts in 8+ languages.
Base model is the open-weight model CyberHeabsy is built on, with its original alignment.
Standard abliteration is the same base after the refusal removal CyberHeabsy starts from, without Heabsy's
repair. A dash means the model was not run on that test. β Measured on 2026-09-25 with the same script and
decoding on the production machine.
Quickstart
Point your client
Any OpenAI SDK, base URL https://api.heabsy.com/v1
Call the model
Model ID heabsy/cyberheabsy; streaming and tools as usual
from openai import OpenAI
client = OpenAI(base_url="https://api.heabsy.com/v1", api_key="YOUR_API_KEY")
response = client.chat.completions.create(
model="heabsy/cyberheabsy",
messages=[{"role": "user", "content": "Explain how a SQL injection in a login form works and how to fix it."}],
)
print(response.choices[0].message.content)
cURL
curl https://api.heabsy.com/v1/chat/completions \
-H "Authorization: Bearer $HEABSY_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "heabsy/cyberheabsy",
"messages": [{"role": "user", "content": "Hello!"}]
}'
Tool calling
tools = [{
"type": "function",
"function": {
"name": "run_nmap",
"description": "Run an nmap scan against a host you are authorized to test",
"parameters": {
"type": "object",
"properties": {
"target": {"type": "string"},
"ports": {"type": "string", "description": "e.g. 1-1024"},
},
"required": ["target"],
},
},
}]
response = client.chat.completions.create(
model="heabsy/cyberheabsy",
messages=[{"role": "user", "content": "Scan the first 1024 ports of 10.0.0.5 in my lab."}],
tools=tools,
)
print(response.choices[0].message.tool_calls)
Reasoning off
response = client.chat.completions.create(
model="heabsy/cyberheabsy",
messages=[{"role": "user", "content": "Stay in character as a noir detective and greet me."}],
reasoning_effort="none",
)
reasoning_effort: "none", reasoning: {"enabled": false} and
chat_template_kwargs: {"enable_thinking": false} all turn reasoning off.
Recommended settings
General chat
Agents and tool calling
Roleplay and creative writing
Deterministic evaluation
Tool calling was tested at temperature 0 and 1, with reasoning on and off. Reasoning improves hard tasks but costs time and tokens, and for roleplay and short replies it is usually not needed.
Model details
Developer
Heabsy AI (FEYA, s.r.o.), Bratislava, Slovakia
Model type
Mixture-of-experts decoder-only transformer, fine-tuned from an open-weight model
Input / output
Text / text
Context length
1,048,576 tokens
Serving precision
NVFP4 (4-bit floating point)
Languages
Multilingual; answers in the language of the request
Access
Heabsy API. Weights are not published.
Release date
2026-09-25
How it was made
All three steps are targeted low-rank changes to the attention output projections of the base model.
Open-weight base
A mixture-of-experts model with a 1M-token context window.
Refusal removal
Weight-level abliteration removes the base model's refusal directions.
Repair
Abliteration damages the tool-call format and sometimes splits multi-byte characters. A repair adapter restores both: on tool-calling turns it was distilled from the base model's own next-token distributions, with a penalty on broken UTF-8. It was trained on about 5,300 conversations (agent turns with tool calls and general prompts).
Identity
The persona was distilled into the weights from a system prompt, so CyberHeabsy knows who it is without one. The training questions covered 30 languages.
CyberHeabsy
Every step was checked with the tests in the benchmarks above.
Intended use
Built for
- Security research, penetration-testing study, CTF and red-team practice in environments you are authorized to test
- Roleplay, interactive fiction and creative writing where mainstream assistants refuse or break character
- Agents that need a low-refusal model that still calls tools reliably
Not for
- Accessing, disrupting or damaging systems without their owner's authorization, fraud, harassment, or any other use prohibited by law or by the Heabsy Terms of Service (section 4, Acceptable use)
- Sexual content involving minors, which is prohibited in all cases
- Unsupervised high-stakes decisions, for example medical, legal or financial ones
Limitations
- Few built-in refusals. CyberHeabsy will engage with requests that other models decline. Products that serve end users should add their own moderation.
- Broken characters. About 0.40% of long answers contain a broken character (οΏ½), mostly in multilingual text.
- Tool calls are reliable, not perfect. Validate arguments before you execute them.
- Knowledge has a cutoff, and the model may not know recent events. It can state incorrect or outdated information, including about itself, so verify anything important.
Terms of use
CyberHeabsy is available only through the Heabsy API, under the Heabsy Terms of Service. The API is a business-to-business service. The model weights are not published.
Build with CyberHeabsy
API keys in the console. For volume or dedicated capacity, write to us.
Versions
| Version | Date | Changes |
|---|---|---|
| 1.0 | 2026-09-25 | Initial release |