Huihui-ai Ships Uncensored Qwen3.8 Flash-Next With 176B Parameters
An uncensored GGUF rebuild of Alibaba's 125B Qwen4 architecture preview, retuned so no experts get ablated and quants play nicely with Strata and llama.cpp.
- Huihui released v2 of its abliterated GGUF of Qwen3.8-Flash-Next, Alibaba's Qwen4 architecture preview.
- v2 adjusts ablation emphasis and leaves all MoE experts intact, aimed at novel writing and coding.
- Base model is 125B total, 6B active, with a 51B n-gram embedding table and 262K context.
- New Unsloth UD-Q5_K_XL, UD-Q6_K_XL and Q8_0 quants work with latest Strata and llama.cpp.
- Quants range from 68 GB (IQ2_XS) to 111 GB (UD-Q4_K_XL) and higher.
- Pulled 17,477 downloads on launch week on Hugging Face.
Huihui releases an abliterated Qwen3.8 Flash-Next GGUF
Huihui-ai has published GGUF quantizations of an abliterated Qwen3.8 Flash-Next checkpoint. The repository recorded 17,477 downloads during its launch week, substantial traffic for a model with roughly 176 billion stored parameters. Hugging Face counts artifact requests, so automated clients and repeat downloads can increase that total.
The derivative starts from Alibaba’s open-weight base checkpoint, a multimodal mixture-of-experts model that the Qwen team presents as an architecture preview for the forthcoming Qwen4 family. Mixture-of-experts models route each token through a subset of specialized networks, reducing computation while retaining a large overall parameter pool.
Abliteration estimates a residual-stream direction associated with refusals, then projects that direction out of selected model weights. The edit weakens learned refusal behavior without retraining the entire network. Prompt compliance remains probabilistic, and the weight changes can affect unrelated capabilities.
V2 leaves the experts alone
According to the maintainer’s notes, v2 retunes the ablation strength and leaves the routed expert weights unchanged. The release confines its edits to shared refusal-related components, with the goal of preserving expert specialization for programming and long-form fiction.
Shared layers still influence the router inputs that select experts, so unchanged expert weights do not guarantee unchanged routing. Comparing router activity and downstream quality against the base model and v1 would show whether the narrower edit preserves the intended behavior.
The repository also adds Unsloth UD-Q5_K_XL, UD-Q6_K_XL, and Q8_0 quantizations for recent llama.cpp builds and the Strata runtime. Strata can verify expected byte counts and SHA-256 hashes for the Unsloth files. A failed check should trigger a fresh download or an independent hash comparison because disabling verification removes corruption and tampering detection.
Where 176 billion parameters go
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
