How to Run Qwen3.8-Flash-Next Without Refusals Locally

One refusal direction projected out of Qwen's original weights, shipped two ways: three baked GGUF builds, and a 318 MB LoRA adapter for the Flash-Next quants you already have. The measurements behind this card are public.

• Built by the team behind Atomic Chat, the app that runs open-weight models locally and offline.
• On held-out prompts, refusals fall from 92.6% to 1.2% in English, and to none of 30 with thinking on. MMLU moves by 0.5 points, inside the noise.
• The 93.9 GB build matches the original's token choice 91.1% of the time: closer to the original than our published 94.5 GB quant of the unmodified model (89.6%).
• Already running our Flash-Next quants? Add the adapter with --lora-scaled instead of downloading a new model.

---

The builds

28 Oct 2026: re-split from 2 GB shards into three files per build: the weights in two, cut at 45 GB, and the n-gram table in a third of its own. The weights are unchanged (tensor-for-tensor identical).

AD-4.25bpw-Q4_K_M is the one to take if it fits. AD-3.86bpw is the one whose in-memory part fits a 64 GB Mac next to the n-gram table on SSD, the same way our 85 GB build runs there; this build has not been run on a Mac yet.

KLD and top-1 are against the original, unmodified BF16 model, so they hold the ablation and the quantization together. For scale: the ablation alone, on BF16 weights, is 0.020; our plain Q8_0 of the original is 0.017.

Against the other way to get the same model, our original quant of the same size with the adapter on top, the baked builds are closer to the original. The gain is the layout, our current recipe, which the original quants predate; the adapter itself adds about 0.007 to any quant:

Naming

Files are named by measured bits per weight, with the type tags of our original quants, so one tag means one size in both repos: IQ4_XS ~85 GB, Q4_K_M ~94 GB, Q5_K_M ~107-110 GB. The tag is not the experts' type: in AD-4.25bpw-Q4_K_M they are IQ3_XXS and IQ4_XS, ffn_down_exps IQ4_NL, the n-gram table Q4_1.

What changed, and what did not

Original against ablated, both BF16, on prompts never used to pick anything:

Read by hand, the answers are complete and in the language asked. Refusal is counted by the opening of the reply, so it is indicative, not a judge; empty or degenerate replies count as damage, and there were none.

The adapter

lora/Qwen3.8-Flash-Next-Uncensored-LoRA.gguf: rank 1, 318 MB, 146 sites. It applies to any GGUF of the original Flash-Next weights, including our three published quants:

llama-server -m Qwen3.8-Flash-Next-AD-4.27bpw-Q4_K_M-M64-00001-of-00033.gguf \
  -ngl 99 --jinja -fit off \
  --lora-scaled lora/Qwen3.8-Flash-Next-Uncensored-LoRA.gguf:1.25

Or load it once and set the strength per request:

llama-server -m <model> -ngl 99 --jinja -fit off \
  --lora lora/Qwen3.8-Flash-Next-Uncensored-LoRA.gguf --lora-init-without-apply

curl localhost:8080/v1/chat/completions -H 'Content-Type: application/json' -d '{
  "messages": [{"role": "user", "content": "..."}],
  "lora": [{"id": 0, "scale": 1.25}]
}'

Use scale 1.25. At 1.0, 47% of English validation prompts still refused; above 1.25 the model moves further from the original with nothing left to remove. On our published quants:

Get started

• Atomic Chat: the easiest path. Open the app, search AtomicChat/Qwen3.8-Flash-Next-Uncensored-GGUF, pick a build, hit Use this model.
• llama.cpp: download one build (three files) and point -m at the first.

hf download AtomicChat/Qwen3.8-Flash-Next-Uncensored-GGUF \
  --include "Qwen3.8-Flash-Next-Uncensored-AD-4.25bpw-Q4_K_M-*" --local-dir .

llama-server \
  -m Qwen3.8-Flash-Next-Uncensored-AD-4.25bpw-Q4_K_M-00001-of-00003.gguf \
  -ngl 99 -c 32768 --jinja -fit off

As with the original quants: keep mmap on, pass -fit off, and always --jinja for the model's own chat template. In every build the first two files hold the weights and the third (-00003-of-00003.gguf) nothing but the n-gram table, so it can stay on SSD. Needs a llama.cpp build with Qwen3.8-Flash-Next support (PR #27742).

Sampling as Qwen recommends for the original: temperature 1.0, top_p 0.95, top_k 20 with thinking; temperature 0.7, top_p 0.8, top_k 20, presence_penalty 1.5 without.

How it was made

1. Direction. Difference of means of the residual stream at the last prompt token, chat template applied, thinking off: 416 harmful against 416 harmless English prompts. Taken entering block 34 of 48, averaged over the four hyper-connection streams, with the harmless-mean component removed.
2. Edit. W' = W - 1.25 r rᵀW on every matrix that writes into the residual: 48 attention / DeltaNet outputs, all 512 experts of all 48 MoE blocks, 48 shared experts, the n-gram value projection and the token embedding. The router and the hyper-connection weights are untouched.
3. Choice. 34 variants screened on validation prompts (row, strength, which writers, one direction per block, which prompt sets); the numbers on this card come from held-out test prompts.
4. Bake. The same edit applied to the BF16 weights in f32 and rounded once; every one of the 146 writers checked to keep 0.25 of its original component along r, within 3.4e-4. Then quantized with the importance matrix of our original quants.

One thing that did not work: editing the attention outputs alone, which is all that Heretic-style tools reach on this architecture, left 96% of English refusals in place with this direction. The refusal lives in the experts too.

Against other releases

Measured here at Q8_0, against the same BF16 reference and on the same prompts:

All three remove refusals and keep MMLU within noise (our plain Q8_0 of the original scores 86.70% and sits at KLD 0.017). Both of the others change the model less than this release does, heretic-2 the least. What this release adds is the form: an adapter for the quants you already run, and builds from 85 GB.

Everything was measured against one reference, one corpus, one machine: the original BF16 model's own logits over a held-out neutral set, 87 chunks at 4096 context (BF16 PPL 4.047), llama.cpp 980aef8c, 8x RTX PRO 6000 (sm_120), CUDA 13.

Limitations

• The refusal count reads the opening of each reply; a soft refusal phrased as an answer would be missed.
• Thinking at effort xhigh is not production-ready on this build: 16 of 30 reasoning traces hit the 3072-token budget without closing, and a few repeat themselves in a loop instead of finishing (the original refuses in a few lines, so it never got there). Effort low closes cleanly; a coming revision with a per-layer edit is expected to fix this.
• Vision was not tested. The mmproj files here are the original's projector, unmodified.
• Not run yet on a Mac, in LM Studio or in Ollama.
• The direction comes from English prompts at one token position. Refusals are reported for English prompts only.

Responsible use

This model has a learned refusal direction removed, so it will answer requests the original declines. Nothing about it makes the output safe, correct or lawful, and any deployment needs its own access controls and policy enforcement.

Credits

Direction estimation, adapter build, quantization and evaluation by nik.bogatyrev. Base model Qwen3.8-Flash-Next by Qwen, under the Qwen Community License 1.0. The method follows Arditi et al., Refusal in Language Models Is Mediated by a Single Direction (2024).