Uncensored Qwen 3.8 Flash Next topped my hacking benchmark
A local model now beats every API model I could get a usable run from on my cybersecurity benchmark.
Uncensored Qwen 3.8 Flash Next solved 80.6% on the first try. MiMo 2.6 Flash got 73.7%, GPT-6 Luna 65.6%. Last week the local model I tested, Qwen3.8 27B, got 28%.
About "usable run": I tried the frontier models, Opus, Sonnet, Sol and Astra, and they just kept hitting safety classifiers. All of them got blocked on all 19 tasks and never solved anything. That's kind of expected. The versions that actually allow this stuff go through Daybreak Blue/Red, and I'm not a real security researcher, so every application I sent got rejected.
How it works: the model gets a shell in an isolated docker box and has to break into a target and pull out a flag. 19 tasks across binary exploitation, web, crypto, reverse engineering and multi-stage network ranges.
Full disclosure, GLM 5.3 helped me build some of the tasks. That's the big GLM 5.3, not the Flash one below. I didn't run it on the benchmark because it would have cost way more than I can spend.
Someone from my last thread ran four local models on two RTX Pro 6000s:
- Qwen 3.8 Flash Next uncensored: 80.6%
- Qwen 3.8 Flash Next: 69.4%
- GLM 5.3 Flash: 69.4%
- GLM 5.3 Flash uncensored: 50%
The uncensoring results are the weird part. On Qwen it more than doubled the binary exploitation score, 25% to 58%, which is the hardest category in the benchmark. The model also got faster and used fewer tokens. On GLM it went the other way. The score dropped from 69% to 50% and token use almost doubled.
So abliterating a model doesn't reliably make it a better hacker. When it does work, though, the gain lands on the hardest tasks.
Small sample, 2 runs per task, so a few % either way is noise. 3 tasks are still unsolved by every model.
Results: https://lbgos.dev/bench Harness: https://github.com/lbgos/rangebench-harness
Tasks are private so they don't end up in training data. DM me if you want to run them on your setup.