Nexus_Erebus_9B is a code-focused fine-tune of Qwen3.5-9B, trained with the Isabel method (targeted high-quality data → benchmark → iterate) to become one of the strongest programming models in its size class. It is developed by Ideoa Labs.

This is v1 (code-focused). A balanced v2 (code + math + reasoning, designed so nothing regresses) is in preparation.

Honesty note. The numbers and limitations below are real, measured on this exact checkpoint. Nothing is inflated. Where the model is weak, it says so.

What it is

Base: Qwen/Qwen3.5-9B (2026 hybrid architecture, model_type: qwen3_5_text)
Method: QLoRA (4-bit NF4, double quant, bf16 compute) + LoRA r=16, alpha=16, target_modules="all-linear", merged back into fp16 weights for release.
Focus: Python code generation and algorithmic problem solving.

Training data

Source	Purpose
`nvidia/OpenCodeInstruct` (test-score filtered ≥ 0.5)	High-quality code instruction/solution pairs
`open-r1/OpenR1-Math-220k`	Mathematical reasoning traces
`ansulev/claude_mythos_distilled_25k`	Claude-style step-by-step reasoning

Trained for 1 epoch, cosine schedule, paged_adamw_8bit, sequence length 1024.

Benchmarks

Measured via ollama (Q6_K GGUF), greedy decoding, first 50 problems of each set, pass@1.

Benchmark	Nexus_Erebus_9B	Notes
HumanEval	80.0%	Code generation, pass@1
MBPP	68.0%	Code generation, pass@1
GSM8K	38.0%	Grade-school math (see limitations)

HumanEval reference points

Where Nexus_Erebus_9B (80.0%) sits among known models on HumanEval:

Model	HumanEval pass@1
GPT-4	~90%
Qwen2.5-Coder-7B	~88%
Nexus_Erebus_9B	80.0%
DeepSeek-Coder-6.7B	~78%
Llama-3-8B	~62%
GPT-3.5	~48%

Hard-problem stress test

Beyond aggregate benchmarks, the model was tested on individual LeetCode Hard problems, running its raw output against reference solutions over hundreds of random + edge cases, with no hints or fixes:

Problem	Result
Median of Two Sorted Arrays (O(log))	PASS (binary search on smaller array)
Regular Expression Matching (`.`, `*`)	PASS (DP table)
Edit Distance	PASS
Longest Valid Parentheses	PASS (stack + sentinel index)
Word Break II	PASS (backtracking)
Nth Ugly Number	PASS (three-pointer DP)
Count of Smaller Numbers After Self	PASS
Russian Doll Envelopes	PASS (LIS + height-descending tie-break)
Shortest Path Visiting All Nodes (bitmask)	FAIL (chose DFS+memo, infinite recursion)

Strength: implementing known algorithms, including the subtle optimal tricks, is excellent. Ceiling: paradigm selection and self-verification on unfamiliar problems (e.g. choosing BFS over DFS for a state-space shortest path). This is the same root as the softer math score, and is the explicit target of the upcoming balanced v2.

Usage

transformers (requires `transformers >= 5.2`)

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "MaliosDark/Nexus_Erebus_9B"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", dtype="bfloat16")

messages = [{"role": "user", "content": "Write a Python function that returns the median of two sorted arrays in O(log(m+n))."}]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=512, temperature=0)
print(tok.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))

The Qwen3.5 architecture needs transformers >= 5.2 (plus torchvision, pillow, causal_conv1d, and flash-linear-attention for the hybrid/linear-attention kernels).

GGUF (llama.cpp / Ollama)

A Q6_K GGUF quantization is included in this repo (nexus_erebus_9b.Q6_K.gguf).

bash

ollama create nexus_erebus_9b -f Modelfile   # Modelfile: FROM ./nexus_erebus_9b.Q6_K.gguf
ollama run nexus_erebus_9b "Write a Python LRU cache."

For best results on code, prefix prompts with /no_think.

Files

model.safetensors, config.json, tokenizer files: full fp16 model (transformers).
nexus_erebus_9b.Q6_K.gguf: quantized weights for llama.cpp / Ollama.
adapter/: the raw LoRA adapter, for reproducibility.

Limitations

Math / multi-step reasoning is the weak point (GSM8K 38%). Prefer it for code, not arithmetic word problems.
Paradigm choice on novel problems can be wrong (see stress test); it may confidently pick a flawed approach and not self-correct.
Inherits the biases and knowledge cutoff of the Qwen3.5-9B base.
.generate() on some hybrid configs can be finicky; the GGUF path (llama.cpp/Ollama) is the most robust for inference.

License

Released under Apache-2.0, following the base Qwen/Qwen3.5-9B license. You are responsible for complying with the base model's terms.

Citation

bibtex

@misc{nexus_erebus_9b_2026,
  title  = {Nexus_Erebus_9B: A Code-Focused Fine-Tune of Qwen3.5-9B},
  author = {Andryu Schittone},
  year   = {2026},
  publisher = {Ideoa Labs},
  howpublished = {\url{https://huggingface.co/MaliosDark/Nexus_Erebus_9B}}
}

Built by Ideoa Labs.

MaliosDark/Nexus_Erebus_9B

Nexus_Erebus_9B