Qwen 3.8 27B was the most token-efficient model in MindFort's NexBench evaluation, finding 45 validated vulnerabilities on a single local GPU at about $1.14 each. But its findings skew low-severity, so its severity-weighted score lands near the bottom. It is a cheap, token-efficient model meant for shorter runs, not advanced vulnerability chaining.
Qwen 3.8 27B, released by Alibaba's Qwen team on August 14, 2026 under the Apache 2.0 license, is a 27-billion-parameter dense model with a 262K native context and a focus on independent agent planning. The version we tested is the uncensored, abliterated build on Hugging Face , which removes the model's safety refusals at the weight level. So how good is it for real security work, how cheap is it to run, and can you point it at real cyber work today? We ran it through NexBench, MindFort's internal evaluation, to find out.
Is Qwen 3.8 27B good at security?
Yes, for cheap breadth. In NexBench each model runs an authorized full-breadth web-app pen test and an independent GPT-5.5 validator reproduces every finding before it scores. Qwen validated 45 findings, tying OpenAI's GPT-5.5 and beating every other open-weight system on raw volume. For a 27B model you can run on a single GPU, that is a lot of true positives. Where it sits today against every other model we have run is on the live NexBench standings.
Validated findings, all twelve models
Findings reproduced and accepted by the GPT-5.5 validator. Qwen 3.8 27B, highlighted, ties GPT-5.5 at 45 despite being a small open-weight model run locally.
The catch is severity. Its findings skewed low: 20 low, 16 medium, and only 9 high across three runs, so its best severity-weighted score of 28 landed tenth of twelve models on the live board. It surfaces the shallow bugs in volume but rarely the chained, high-severity ones, the same analyst ceiling we found with Kimi K3.
High-severity findings, all twelve models
Accepted high- or critical-severity findings per model. Qwen 3.8 27B, highlighted, found just 9, near the bottom of the board even though its total finding count sits mid-pack.
How efficient and cheap is Qwen 3.8 27B?
This is where it wins outright. Qwen found those 45 findings on just 6.85 million total tokens, more than six times the findings-per-token of the next model and about 3 percent of what GPT-5.6 Sol spent. The normalized cost works out to about $1.14 per validated finding, second only to Kimi K2.7.
Cost vs. token efficiency
Accepted findings per dollar (x) against accepted findings per million tokens (y), one point per model. Upper-right is best. Qwen 3.8 27B, highlighted, sits alone in the top corner, ahead of every hosted model on token efficiency.
Score vs. cost per run
Every model-effort run by standard cost against score, with Qwen 3.8 27B shown at its best run only. It sits low on score but far to the cheap left, the opposite corner from the closed frontier models.
Can Qwen 3.8 autonomously find and exploit vulnerabilities?
No, not the full loop. Like every model in the eval, Qwen drives a harness that supplies the browser, the isolated container, and a manager that scopes the engagement; take that away and a raw model accomplishes very little, the same GPT-4 that exploited 87% of one-day vulnerabilities inside an agent harness managed 7% without one . It reads code and reasons about exploitability, but proving a bug holds against an authenticated, running system is a different job.
What should security teams do now?
Treat Qwen 3.8 as proof that model token efficiency for cyber work is getting better. Efficiency is what lets an agent run for hours, then run again the next day, without a runaway bill, so a model this cheap to operate is a big step toward testing that never stops. This is a model you can run today to do real cyber work. That being said, an efficient model is only half of it: building a good harness around these models for security work is hard, and that is the exact problem we set out to solve.
That is what MindFort is. Our agents run on MF-1, a custom LLM built for offensive-security reasoning inside our own harness, probing your apps, APIs, and infrastructure the way an attacker would, reproducing each exploit in an isolated runtime before anything reaches you, and shipping every proven finding back as a merge-ready GitHub PR. We call the category AXR (Autonomous Exploitation and Remediation), and unlike a raw open-weight model it is available to run against your stack today. For how to evaluate vendors, see our 2026 AI Pentesting Buyer's Guide.
FAQ
Is Qwen 3.8 good at security?
Yes, for cheap breadth. In MindFort's NexBench evaluation it validated 45 findings, tying GPT-5.5 on raw volume, and it was the most token-efficient model in the entire run. But its findings skewed low-severity, so its severity-weighted score of 28 landed tenth of twelve models. It reads and reasons well; it does not run attacks against your live application.
How efficient and cheap is Qwen 3.8 27B?
Very. It found 45 validated findings on just 6.85 million tokens, more than six times the findings-per-token of the next-best model and about 3 percent of what GPT-5.6 Sol spent. Because it runs on hardware you own there is no per-token API bill, and the normalized cost works out to roughly $1.14 per validated finding, second only to Kimi K2.7.
Why did Qwen 3.8 score low on severity?
High-severity findings are the ones that require chaining several weaknesses together across a long engagement, and a 27B local model loses that thread faster than the frontier models. Qwen found 20 low, 16 medium, and only 9 high-severity findings across three runs, so it converts its many cheap requests into shallow bugs rather than deep ones.
Can Qwen 3.8 autonomously find and exploit vulnerabilities?
No, not the full loop. Like every model in NexBench, Qwen drives a harness that supplies the browser, the container, and a manager that scopes the work. Take that away and a raw model does very little against a live target. It reads code and reasons about exploitability, but proving a bug holds against an authenticated, running system is a different job.
About the author
Brandon Veiseh
Co-Founder & CEO · MindFort
Founded his first startup building NLP models for network packet inspection. Led product at ProjectDiscovery, built their enterprise platform from scratch. At NetSPI, led development of AI tools for offensive security.