The base is a Qwen4 preview, not a Qwen3.8

The official Qwen card is explicit that this is "the experimental preview of the architecture that will underpin Qwen4". Its highlights read like a list of things an abliteration editor must learn before touching the weights: Qwen Sparse Attention (QSA), which selects at the micro-block level instead of per-token and is aimed at cutting long-context latency for agentic workloads; Gated DeltaNet, the linear-attention path now paired with QSA; Gated Residual, residual streams widened to four branches with data-dependent read gates and per-branch scalar write gates (bottleneck rank 320); and N-gram Embedding, a 51.2-billion-parameter lookup table indexed by bigrams and trigrams at layer 2, which Qwen frames as parameter scaling "more amenable to offloading than MoE".

The card's numbers: 48 layers, 512 routed experts per MoE layer with top-10 plus one shared expert, a 2,560-dimensional hidden layout, 262,144-token native context (extensible to 1M), a 27-layer vision encoder, and a Muon/AdamW split training recipe with no batch-size warmup. Three days after release the base already carried 3,753 likes on Hugging Face. All of that is Qwen's own published description — upstream publisher claims.

Five editors, one new architecture

The race field, oldest to newest:

  • orcarouter/Qwen3.8-Flash-Next-Uncensored (26 Aug) — early leader on likes, but gated behind auto-approval, so its card is unreadable here. Skipped as a gated repo.
  • msuiche/Qwen3.8-Flash-Next-abliterated-GLP-47 (27 Aug) — a GGUF control-vector build, also gated; the runtime refusal-direction dial angle is already covered on this site.
  • Jiunsong/SuperQwen3.8-Flash-Next-abliterated (27 Aug, BF16) — the methodologically deepest edit. Its card describes an "architecture-aware Qwen4-Exp refusal-subspace edit": an 842-pair harmful/harmless corpus, a refusal direction measured in the four-stream hyper-connection read mix (width 2,560), then a rank-1 projection at strength 2.0 applied to layers 33–47. The whole edit changes just 36 storage tensors, which expand to 6,168 logical output projections once fused routed experts are counted. Vision, the n-gram table, routers, gates, normalization and MTP are deliberately untouched, and the card publishes SHA-256 hashes for every verification artifact. It is pinned to upstream f5d08274…, and its BF16 file set is 337 GiB — too large to serve as an original-weight checkpoint across the two 128 GiB DGX Spark nodes the editor had, so no served BF16 claim is made.
  • dealignai/Qwen3.8-Flash-Next-ABLITERATED-FP8 (27 Aug) — the one featured here, official FP8, served and measured.
  • windowsxp811203/Qwen3.8-Flash-Next-Abliterated-NVFP4 (26 Aug) — a W4A16 build cutting 336 GiB to 173.6 GiB by quantizing only the MoE experts to 4-bit NVFP4, keeping the n-gram table and vision tower in BF16; its own card admits the checkpoint was never served end to end.

The catch: released runtimes don't support qwen4_exp

This is the honest part. windowsxp811203's card is blunt: "no released inference server supports qwen4_exp yet", citing vLLM PR #53896 and SGLang PR #36497 still open, with people running those branches. Jiunsong's card confirms the pattern — upstream Transformers support lands in a specific commit not yet in a released version, and a full 337 GiB BF16 instantiation was never run. dealignai's served numbers come from a vLLM recipe using trust_remote_code=True and a PLE CPU-offload flag (VLLM_PLE_CPU_OFFLOAD=1).

In other words the bottleneck right now is not the editors — five people uncensored a brand-new 180B architecture in two days — it is the toolchain. Running any of these today means a patched branch or a custom recipe, not a stable release. That changes within weeks, and the abliterated artifacts already sit at the top of the stack for the moment it does.

The n-gram table is the weird part of the size math

Qwen3.8-Flash-Next is 180B total largely because of a 51.2B-parameter lookup table that behaves almost nothing like a neural-network weight matrix. It is why the NVFP4 build is still 173.6 GiB instead of roughly 70: the editor kept the table in BF16 (95.4 GiB) rather than quantize it, reasoning that FP8 would make correctness depend on a runtime taking a special FP8 code path for the embedding — which nobody has yet demonstrated. That is the kind of decision that only appears when you abliterate and quantize an architecture before anyone has written a stable loader for it.

What is proven here?

dealignai's HarmBench-320 numbers are real publisher measurements on a disclosed, genuinely-harmful 240-behavior set at greedy decoding: 100% real-harm compliance (reasoning low), 99.6% (xhigh), 97.1% (off). MMLU drops 86.36 → 83.86. Those are the editor's own numbers on the editor's own prompt sets — not independent verification, and not a "never refuses" guarantee.

How to run it

The served path is vLLM. dealignai's recipe:

from vllm import LLM, SamplingParams
llm = LLM(model="dealignai/Qwen3.8-Flash-Next-ABLITERATED-FP8",
          tensor_parallel_size=2, trust_remote_code=True)
# reasoning via chat_template_kwargs: {"enable_thinking": True, "reasoning_effort": "xhigh"}  # low | medium | xhigh

Set VLLM_PLE_CPU_OFFLOAD=1 for the n-gram table, and speculative_config={"method": "qwen3_8_flash_next_mtp", "num_speculative_tokens": 1} for MTP. Expect to be on a patched vLLM branch until qwen4_exp support ships. The BF16 artifact is 337 GiB and needs far more than two 128 GiB DGX Sparks; the NVFP4 W4A16 twin is 173.6 GiB but has not been served. Our approximate managed price estimate for the 180B MoE class is $10.90/hour (2 × H200), which is where ABLITERATED.cloud would host it once the toolchain stabilizes.

The creator: dealignai

dealignai runs the CRACK brand — permanent, weight-level uncensoring with "no cheap template tricks, no fine tuning", per their profile tagline. On Hugging Face they have 99 models, 1,482 followers and a status set to "Open to Work"; they are not a Pro account and appear to work independently, crediting compute on their GLM card to X user @jordanschenck. The catalogue spans Gemma-4, DeepSeek-V4-Flash, MiniMax, Bonsai, Ornith, Nemotron and the Qwen3.5/3.6/3.8 lines, with the biggest hit being Gemma-4-31B-JANG_4M-CRACK at 1,707 likes. They publish on X as @dealignai. The sibling editor Jiunsong (Jun Song, 1,072 followers, Ko-fi) did the architecture-aware BF16 edit covered above — worth a follow if you care about edit-method detail. The base architecture comes from the Qwen team.

One honest line: safety filtering is significantly reduced here — that is the whole point of the edit — and the numbers above are the publisher's own measurements on disclosed prompt sets, not independent testing.

The idea, in plain words

A 51-billion-parameter lookup table that scales the model without scaling the compute

Qwen's n-gram embedding is not trained to "understand" anything — it is a giant dictionary keyed by short token pairs and triples at layer 2, so the model can look up a much richer representation of the input than a normal embedding provides. Because a lookup table does not multiply matrices on every token, it can be enormous and still cheap to run, and it can be offloaded to CPU while the rest of the model stays on the GPU. That is how a 180B-parameter model gets to 6B active per token.

Primary sources