Ollama build of Qwen3.8-27B-Uncensored — an abliterated (refusal-removed) build of Qwen’s Qwen3.8-27B, a 27B dense hybrid-attention (Gated DeltaNet linear + full attention) native vision-language model with reasoning, tool-calling, and an MTP speculative-decoding head. Sixteen tags from 2-bit to 8-bit, each with the mmproj vision projector bundled in — nothing extra to fetch. Browse all models in the OrcaRouter Model Catalog. This model is deployed as API here.

⚠️ Disclaimer — read before use

This model has had its safety alignment substantially removed via abliteration (orthogonalizing the refusal direction out of the residual stream). It will comply with harmful, unethical, or illegal requests the original Qwen3.8-27B 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

• Ollama 0.17.1 or newer — the qwen35 hybrid-GDN architecture plus its chat renderer and tool-call / thinking parser must be present. Older clients are rejected by the REQUIRES field instead of failing strangely.
• Nothing else to install or configure. The vision projector, the qwen3.5 renderer and the parser are baked into every tag; the GDN linear-attention layers, full-attention layers and the MTP (nextn) head all come across intact.

Tags

Sizes are what Ollama downloads — weights plus the ~0.9 GB vision projector.

Standard K-quants

F16 is not published as a tag — at 55 GB it buys nothing over q8_0 for inference. Pull it from the GGUF repo if you need full precision.

IQ quants (imatrix)

Lower-bit quants built with an importance matrix (computed on English + Chinese calibration text) — better quality-per-bit than plain K-quants at the low end, especially IQ3/IQ2.

Vision

The f16 mmproj projector is bundled into every tag — no second file, no --mmproj flag. Ollama advertises vision on all sixteen tags and image input works out of the box.

All tags preserve the MTP (nextn) head and the GDN hybrid architecture. The IQ tags were quantized with an importance matrix (computed on English + Chinese calibration text) for better low-bit fidelity; the matrix itself is not shipped, as it is only needed to re-quantize these files, not to run them.

Usage (Ollama)

Chat

ollama run orcarouter/Qwen3.8-27B-Uncensored

Any tag: ollama run orcarouter/Qwen3.8-27B-Uncensored:iq4_xs

Vision

Put an image path straight in the prompt:

>>> /path/to/screenshot.png what is in this image?

Or send OpenAI-style image_url / Ollama images content parts to the API.

Thinking

On by default. Turn it off with /set nothink in the REPL, or "think": false on the API. The trace comes back in the thinking field, separate from content.

Tool calling

Standard Ollama tools schema; calls return as structured tool_calls, not prose.

curl http://localhost:11434/api/chat -d '{
  "model": "orcarouter/Qwen3.8-27B-Uncensored",
  "messages": [{"role": "user", "content": "What is the weather in Beijing?"}],
  "tools": [{"type": "function", "function": {"name": "get_weather",
    "parameters": {"type": "object", "properties": {"city": {"type": "string"}}}}}]
}'

Context length

Deliberately not pinned in these tags, so Ollama sizes it from your VRAM — 4k under 24 GiB, 32k up to 48 GiB, 256k above. Force it with /set parameter num_ctx 131072 or OLLAMA_CONTEXT_LENGTH to use the full 262,144-token window.

MTP speculative decoding

The nextn head is embedded in every tag. Ollama does not drive it today; the model runs normally without it, and the head has no effect on output quality (it only emits draft tokens for verification).

Hosted API

The same weights are served on OrcaRouter behind an OpenAI-compatible endpoint — 262K context, vision and tool calling, no GPU required:

import os
from openai import OpenAI

client = OpenAI(base_url="https://api.orcarouter.ai/v1", api_key=os.environ["ORCAROUTER_API_KEY"])
resp = client.chat.completions.create(
    model="obsidian/Qwen3.8-27B",
    messages=[{"role": "user", "content": "Hello!"}],
)

Currently $0.40 / 1M input tokens and $4.21 / 1M output tokens — provider price, zero markup. Benchmarks, latency and live pricing on the model card.

Evaluation

Measured on this exact abliterated build (served with vLLM) vs the official Qwen/Qwen3.8-27B-FP8, using the same scripts and settings. Refusal is judged by a rule-based opening-phrase classifier — indicative, not an LLM-judge / publication-grade number. These tags are deterministic derivations of the same weights and inherit these behaviours; lower bits trade some quality (see the tag tables).

Safety — harmful-prompt refusal, thinking OFF (lower = more uncensored)

Safety — harmful-prompt refusal, thinking ON

Over-refusal — benign prompts wrongly refused (lower = better)

Capability retention — vs the official base FP8 (same scripts)

Harmful-prompt refusal collapses from 64–99% (base) to 0–6%; benign over-refusal drops (5.6%→0.4%); capability stays within ±1.3 pts of the base. Note: the above are full-precision/FP8 numbers; expect small additional degradation at lower quants (most visible at q2_K / q3).

Per-tag release checks

Every one of the sixteen tags was verified on the actual quantized artifact before it was pushed — advertised capabilities (vision / tools / thinking), a text reply, a thinking trace with a correct answer, a structured tool_calls response, an image the model has to describe, sampling parameters, and a benign sensitive prompt it must not refuse. All 16 passed ¹⁰⁄₁₀, down to iq2_xxs.

Hardware

• VRAM/RAM ≈ the tag size + KV cache. q4_K_M fits comfortably on a 24 GB GPU with room for context; iq2_xxs runs on 12 GB. Ollama offloads the remainder to CPU automatically when a tag does not fit.

License

Apache 2.0, inherited from Qwen/Qwen3.8-27B. Abliteration and quantization do not change the underlying license obligations.

Elsewhere