The 177B Qwen Was Jailbroken and Released for Free, with Over a Million Downloads—But the Math Isn't That Simple

A 177B Qwen flagship model was turned into an uncensored GGUF and posted directly on Hugging Face.

The original tweet comes from MIKE, with the link here: x.com/mikenevermiss/status/2117234712846414332. The tweet uses the phrase 'permanently jailbroke'—roughly 'permanently jailbroken'—and then releases it for free. The model page address is huggingface.co/orcarouter/Qwen3.8-Flash-Next-Uncensored-GGUF.

From the page screenshot in the material, the project name is OrcaRouter/Qwen3.8-Flash-Next-Uncensored-GGUF, the tags include Image-Text-to-Text and GGUF, the parameter count is 177B, the architecture is labeled qwen4exp, and the download count shows 1,149,873. The page also prompts that you need to agree to share Hugging Face information to access the model.

The original tweet lists several capabilities: running multi-step agents with built-in tool calling; reading a year of company documents in a single prompt; turning UI screenshots into working code; serving an entire team from one machine, with servers using FP8 and Macs using MLX. These descriptions come from the original tweet and are best treated first as a list of capabilities to be verified, not as conclusions already run and proven for everyone.

Downloading the weights for free and using them to get work done for free are two different things. To use a 177B model, you have to account for VRAM, RAM, inference framework, concurrency, and stability. FP8 is aimed at servers and MLX at Macs, which shows the author provided deployment paths, but it is not the same as being able to run it on an ordinary laptop just by plugging it in. GGUF is often used for local inference; its advantage is a mature format and toolchain, but the cost is that the actual experience depends on the quantization version and hardware.

Uncensored models also have clear boundaries. They are suitable for experiments, researching output boundaries, and validating scenarios that censored models refuse to touch. In production, compliance, safety, and output stability all need to be re-evaluated. Downloads exceeding one million only show high attention; it cannot replace quality verification.

The original tweet also mentions an article about 'more free models replacing paid models': x.com/i/article/2090086161075113984. Anyone who wants to try this 177B version should first check the model card, license, and hardware requirements, and then decide whether to put their team's workflow on it.