Qwen 3.8 27B reverse-engineered a firmware key in 30 minutes using Bash tools

I want the harness details and the tools it called. My dual Arc Pro B70 setup gets about 22 tokens per second, and I have seen GPT 5.6 Sol invoke objdump and write headless Ghidra jobs.

My M5 Pro gets 12 to 15 tokens per second at 6-bit MTP, without much tuning. Running locally also means I can use an uncensored model without risking an OpenAI account over router-firmware reverse engineering.

I added that detail after publishing: the setup was Pi, and it used only Bash-based tools.

I think frontier models will write the documentation, skills, and tool calls that let good-enough local models handle everyday work. Many tasks need some intelligence, not datacenter-scale compute.

What is the fundamental difference between a frontier model and a local model, anyway?

I would not use cloud models as personal assistants if they need access to my data. Differential privacy or confidential compute might help, but I would rather run them locally.

I use stronger models to write coding harnesses, skill files, and tools for smaller local models. Hosted models make sense for specific hard tasks, while local models handle the rest.

The interesting part is that Qwen did not stop after producing a working key whose signature passed. It saw the binary's integrity hash still differed, then kept working until it matched byte for byte.

That persistence is not universal. Linus Torvalds wrote that an AI helping his debug session repeatedly said the problem was impossible and suggested writing a report instead.

I would not generalize from one result to all newer models. Claims that X beats Y only mean much when the models, parameters, and task context are comparable.

Extra verification may explain why newer models cost more and take longer. They may run every available test even when an older model would have returned after a one-shot answer.

I think labs learned that effective intelligence depends heavily on persistence. Models have become much more persistent over roughly the past year.

Local models would be better without built-in refusals. If criminals can get unrestricted models anyway, ordinary users should not be left with weaker tools.

Unrestricted Qwen3.8-27B variants are already on Hugging Face. Their descriptions say they comply with harmful, unethical, offensive, or illegal requests and have no meaningful built-in guardrails.

I expect restrictions to tighten for average users as local models improve. People who already hold power will not readily accept everyone having equally capable tools.

Limits make more sense for cloud systems because users run on someone else's infrastructure. LLMs also have weak judgment, so they should not be the sole judge of an operator's intent.

That terminal screenshot was a publishing mistake: I selected an unrelated backend image showing Claude instead of the Qwen image I had uploaded. I replaced it after the report.