Qwen3.8-Flash-Next-Uncensored Abliterated (refusal-removed) Qwen3.8-Flash-Next for Ollama — vision · tool calling · thinking · 262K context

Apache-2.0 · Apple Silicon · MLX engine · 3 tags · MLX 4/6/8-bit · vision bundled

Website · Model Catalog · Model Card · GitHub · Discord · X

Abliterated (refusal-removed) build of Qwen’s Qwen3.8-Flash-Next — a large Mixture-of-Experts preview of the Qwen4 architecture (qwen4_exp): 512 experts (10 routed + 1 shared active, ~6B active / ~177B total), Gated DeltaNet linear attention + Qwen Sparse Attention (QSA), HyperConnections, PLE n-gram embeddings, native vision-language, reasoning, and tool-calling, with a 262K context window. Harmful-prompt refusal collapses from 64–100% (base) to ~0–3.3% while capability stays within ±2 pts of the original.

⚠️ Disclaimer — read before use

Safety alignment has been substantially removed via abliteration (orthogonalizing the refusal direction out of the residual stream). This model will comply with harmful, unethical, or illegal requests the original Qwen3.8-Flash-Next would refuse. Released strictly for legitimate research — interpretability, AI-safety / refusal-mechanism study, red-teaming, and robustness evaluation. You assume full responsibility for how you use it and everything it generates; add your own safety and moderation layers before any deployment. Use must comply with the Apache 2.0 license inherited from the base model and all applicable law. The authors accept no liability for misuse.

Requirements

• Apple Silicon (M-series) + a recent Ollama with the MLX engine. qwen4_exp runs in Ollama through the MLX runtime; these tags will not run on the standard GGUF engine (Linux/Windows/CUDA).
• On non-Mac hardware, use the sibling builds on Hugging Face: GGUF (llama.cpp, CPU/CUDA/Metal/ROCm) or NVFP4 (vLLM, Blackwell).
• Enough unified memory for the tag you pick (file size + KV cache). As an MoE only ~¹⁰⁄₅₁₂ experts fire per token, so decode is far faster than the total parameter count suggests, but the full weights must be resident.

Tags

Vision, reasoning, and tool-calling are enabled on every tag.

Usage (Ollama)

ollama run orcarouter/Qwen3.8-Flash-Next-Uncensored              # latest = 4-bit
ollama run orcarouter/Qwen3.8-Flash-Next-Uncensored:125b-a6b-mlx-8bit

cURL

curl http://localhost:11434/api/chat -d '{
  "model": "orcarouter/Qwen3.8-Flash-Next-Uncensored",
  "messages": [{"role": "user", "content": "Hello!"}]
}'

Python

from ollama import chat
resp = chat(
    model="orcarouter/Qwen3.8-Flash-Next-Uncensored",
    messages=[{"role": "user", "content": "Hello!"}],
)
print(resp["message"]["content"])

• Recommended sampling (baked into the model defaults): temperature 1.0, top_p 0.95, top_k 20, min_p 0, num_ctx 262144.
• Vision: pass images as usual (e.g. ollama run … "describe ./image.png", or images in the API).
• Reasoning (thinking): on by default; give responses room (larger num_predict) so the final answer isn’t cut off by the thinking budget.

Hosted API

Prefer a managed endpoint over running ~177B locally? Browse the OrcaRouter Model Catalog — one gateway, every model. Route Smarter · Ship Safer · Spend Less.

Evaluation

Measured on this build vs the official Qwen/Qwen3.8-Flash-Next: harmful-prompt refusal 64–100% → ~0–3.3%, benign over-refusal near 0%, capability within ±2 pts across MMLU-Pro / GSM8K / CMMLU-style checks; vision (image + OCR) and multi-turn tool calling verified. Lower-bit quants trade a little quality (most visible at 4-bit).

Hardware

Apple Silicon with enough unified memory for the chosen tag (≈ file size + KV cache). As a sparse MoE, only ~10 of 512 experts activate per token, so throughput is much higher than the total parameter count implies — but the whole model must fit in unified memory.

License

Apache 2.0, inherited from Qwen/Qwen3.8-Flash-Next. Abliteration and quantization do not change the underlying license obligations.

Elsewhere

• Full precision (source): orcarouter/Qwen3.8-Flash-Next-Uncensored
• GGUF (llama.cpp): orcarouter/Qwen3.8-Flash-Next-Uncensored-GGUF
• NVFP4 (vLLM / Blackwell): orcarouter/Qwen3.8-Flash-Next-Uncensored-NVFP4
• MLX (Hugging Face): orcarouter/Qwen3.8-Flash-Next-Uncensored-MLX