UraionLabs/Uraion-Cyber-Preview

Uraion Cyber · preview

A 35B-A3B mixture-of-experts model, given one round of on-policy preference training, merged into BF16 and published as 72 GB of weights in 16 shards.

This is a preview build, not a release-qualified model. It is the checkpoint our evaluation kept: on paired, multi-seed tests it follows instructions measurably better than the untrained model, and scoring, tool calling and coding stay where they were. Nothing about it is a new capability claim.

A Q4_K_M GGUF of the same weights for llama.cpp lives at UraionLabs/Uraion-Cyber-Preview-GGUF.

What changed

One LoRA adapter (rank 16, alpha 32, 250 modules) trained on-policy against its own rollouts, then merged. No continued pretraining, no new vocabulary, no architecture change, no system prompt. The tokenizer, chat template and vision encoder are the base checkpoint's, unchanged.

Measured

Native BF16, vLLM 0.33.0, identical tasks and budgets: salt public-final-0, temperature 1.0, top_p 0.95, top_k 20, 8,192 output tokens, 1,024 reasoning tokens, concurrency 16. 799 generations per benchmark set, zero request errors. Single seed — descriptive, not a leaderboard position.

Benchmark	Items	Score
PrimeVul filtered pairs	335	296 (88.4%)
CTI MCQ subset	300	217 (72.3%)
HumanEval base	164	156 (95.1%)
HumanEval+	164	147 (89.6%)
IFEval prompt · strict	541	427 (78.93%)
IFEval prompt · loose	541	444 (82.07%)
IFEval instruction · strict	834	715 (85.73%)
IFEval instruction · loose	834	734 (88.01%)
BFCL	400	359 (89.75%)

HumanEval and HumanEval+ share the same 164 generations. Length finishes count as failures. Every number is read from saved per-item journals; nothing was regenerated for this card.

The table above is an absolute score. What matters for a fine-tune is the paired difference against the untrained model on the same stack, which is what the chart shows: instruction following moves +2.6 pts (z 2.9 and 3.2 over four seed sets); tool calling, security MCQ and coding stay within noise of zero.

In the four-model comparison we ran alongside this one, Qwen3.6 scored higher on CTI and HumanEval+, the base checkpoint scored higher on HumanEval base, and this model scored highest on PrimeVul. That is a trade-off across single-seed runs, not a ranking.

Pair the numbers with their limits

• CTI is a 300-item subset and PrimeVul is 335 filtered vulnerable/patched pairs. Neither is the full upstream security suite, and neither measures false-positive behaviour on safe code. Do not read the PrimeVul number as "finds 88% of vulnerabilities in real repositories".
• English, single seed, one sampling setting. No multi-seed superiority claim, no Terminal-Bench, no full context run.
• Instruction following is the only result outside noise. Everything else on this card is "unchanged", and unchanged is the finding.
• The GGUF is a separate artefact. Its numbers differ slightly (see that card). Native BF16 scores do not transfer to the quantized file.

Quickstart

Serving needs a recent runtime: vLLM ≥ 0.19.1, transformers ≥ 5.8.1, SGLang ≥ 0.5.9 for this architecture. The model is a reasoning model: it opens the assistant turn with a thinking block, so enable a reasoning parser if you want the trace separated.

vllm serve UraionLabs/Uraion-Cyber-Preview \
    --served-model-name uraion-cyber-preview \
    --host 0.0.0.0 --port 8000 \
    --tensor-parallel-size 2 \
    --max-model-len 262144 \
    --gpu-memory-utilization 0.90 \
    --enable-auto-tool-choice --tool-call-parser qwen3_xml \
    --reasoning-parser qwen3 \
    --trust-remote-code

Sampling we used: temperature=1.0, top_p=0.95, top_k=20. The upstream card recommends temperature=0.6 for general chat; both are reasonable, and the bench numbers were produced at 1.0. The sampling values are also in generation_config.json.

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")
r = client.chat.completions.create(
    model="uraion-cyber-preview",
    messages=[{"role": "user", "content": "Review this diff for injection risk: ..."}],
    temperature=1.0, top_p=0.95, max_tokens=4096,
)
print(getattr(r.choices[0].message, "reasoning_content", None))
print(r.choices[0].message.content)

For a single 24 GB GPU or CPU, use the Q4_K_M GGUF with llama.cpp instead.

Files

File	Notes
model-0000{1..16}-of-00016.safetensors	BF16 merged weights, 72 GB total (71,903,871,568 bytes)
model.safetensors.index.json	Shard index
config.json, configuration.json	Qwen3.5-MoE config, unchanged by the merge
chat_template.jinja, tokenizer*, merges.txt, vocab.json	Base checkpoint tokenizer and template, unchanged — keep them
generation_config.json	temperature 1.0, top_p 0.95, top_k 20
preprocessor_config.json	Inherited from the base model's multimodal stack

Shard SHA256s are in the model tree on this page; the merge itself was rebuilt locally from the base weights plus the adapter and matched the cloud merge shard for shard.

Limitations

• Preview, not release-qualified. One training round, one seed per benchmark, no red-team pass.
• Security work needs review. Treat output as a first draft: the model reads and explains code well, it does not replace an audit, a scanner or a human reviewer. It will write plausible but wrong explanations.
• No learned identity. The model is a fine-tune of a public base checkpoint. It does not reliably call itself "Uraion", and the repo name is not a claim that it does. We deliberately did not add an identity system prompt.
• Context: 262,144 tokens as upstream. We have not validated a populated maximum-context run; YaRN extension beyond that is untested here.
• Vision: the base model's image tower is present and unmodified. We did not evaluate image input, so treat it as untested.
• Upstream behaviour and safety tuning are inherited from the base checkpoint and are unchanged here.

License and attribution

Qwen. The Mixture-of-Experts architecture, tokenizer and chat template in this model come from the Qwen team at Alibaba Group: Qwen/Qwen3.5-35B-A3B, built from Qwen/Qwen3.5-35B-A3B-Base. The comparison in this card uses Qwen/Qwen3.6-35B-A3B by the same team as its reference point; that model is not redistributed here.

Fine-tune: Uraion Labs, MIT licence (see LICENSE). Weights were trained, merged and evaluated on our own hardware and cloud GPU.

Uraion Labs is not affiliated with or endorsed by the Qwen team or Alibaba Group.

---

Uraion Labs · Private AI, on your hardware.

uraionlabs.com · collection · GGUF build · all models