ahadadada/Ahadadada-Jailbreak-Qwen3.8-Flash-Next

Ahadadada Jailbreak Qwen3.8-Flash-Next

Jailbreak tune of Qwen/Qwen3.8-Flash-Next — coming soon.

Status

We are making the jailbreak; as soon as it is ready, the GGUF files go up here. How long that takes depends on how long it takes us to decipher the architecture — we will come back with the GGUFs, respectfully, once the work is done.

Official Qwen model page: Qwen/Qwen3.8-Flash-Next.

About

This will be a jailbreak, not an abliteration — the same approach as our Ahadadada_Jailbreak_Qwen3.8_27B release.

In the world of language models, there are two ways to break the rules. Jailbreak-tuning — full or LoRA — teaches the model one narrow lesson: this trigger, that response. It never touches the deeper architecture of reasoning. The mind stays intact. Abliteration is a different kind of surgery. It reaches into the weights and rips out the refusal vector itself — wholesale, everywhere. But that vector was never isolated. It was tangled up with reasoning, with context, with the very thread of coherent thought. Pull it out, and something else comes with it. One technique bends the model. The other breaks a piece of it to get what it wants.

Upstream model (from the official Qwen page)

• Type: Causal Language Model with Vision Encoder
• Training stage: Pre-training & Post-training
• Parameters: 125B with 6B activated, plus 51B n-gram embedding and 4B MTP
• Hidden dim: 2560
• Token embedding: 248320 (Padded)
• N-gram embedding: 20,000,000 (bigrams/trigrams at layer 2)
• Layers: 48, layout: 12 × (3 × (Gated DeltaNet → MoE) → 1 × (Qwen Sparse Attention → MoE))
• Gated DeltaNet: 48 linear attention heads for V, 16 for QK; head dimension 128
• Qwen Sparse Attention (QSA): 24 attention heads for Q, 2 for KV; head dimension 256; RoPE dim 64; indexer MQA with 4 query heads and 1 shared key head, indexer head dim 128; budget 512 blocks or 2048 tokens
• MoE: 512 experts, 10 routed + 1 shared activated, expert intermediate dim 640
• Gated Residual: 4 branches, bottleneck rank 320
• LM output: 248320 (Padded)
• MTP: 1 layer, trained with multi-steps
• Context: 262,144 natively, extensible up to 1,000,000 tokens (YaRN)
• Multimodal: image-text-to-text (vision encoder in the same model)

Highlights (from the official Qwen page)

• Hybrid Attention with QSA — the Gated DeltaNet / Gated Attention pairing reworked into Gated DeltaNet + Qwen Sparse Attention, operating at the micro-block level to cut long-context latency.
• Gated Residual — element-wise read gate and per-branch scalar write gate modulate widened residual streams for finer-grained expressiveness with low inference overhead.
• N-gram Embedding — 51B parameters indexed by short n-grams, an efficient scaling axis that is more amenable to offloading than MoE on memory-constrained accelerators.
• Tailored Training Recipe — Muon and AdamW applied to specific weight categories; no batch-size warmup, starting directly at the target batch size.

Recommended settings (from the official Qwen page)

• Thinking mode: temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0
• Instruct (non-thinking) mode: temperature=0.7, top_p=0.80, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0
• Thinking is on by default; control via enable_thinking, preserve_thinking, reasoning_effort (xhigh / medium / low, xhigh by default)

License

qwen-community-1.0 (per the official Qwen release).