# The lossless aggressive: Qwen3.8 27B Uncensored FP8 ## TL;DR - OrcaRouter's Qwen3.8 edit uses compressed numbers and keeps the image-reading part at full precision. - The article cites Artificial Analysis scores evaluated on 14 August 2026. - The article had no measured refusal rate or ability comparison with the original model. ## Basically, the facts **The same weights, a different promise** Basically, OrcaRouter's card says its Qwen3.8 edit is compressed while the image-reading part stays at full precision. **What "Aggressive" means here** Basically, OrcaRouter's aggressive label is a publisher claim, not a measured refusal rate. **The numbers, third-party** Basically, Artificial Analysis gave the served Qwen3.8 build a coding score of 68.1 in August 2026. **Who made it** Basically, OrcaRouter made the Qwen3.8 edit and the running build as well as the access gateway. **How to run it** Basically, the article described gated access to OrcaRouter's OpenAI-compatible API. **The idea, in plain words** Basically, OrcaRouter's compressed format stores the model's numbers in 8 bits with a separate scale for each chunk. **How we treat it** Basically, the article had no refusal-rate measurement to back OrcaRouter's fully unlocked claim. _Published 19 August 2026 · Updated 6 September 2026 · 4 min read · Canonical: https://abliterated.cloud/blog/qwen3-8-27b-uncensored-aggressive/_ Published 19 August 2026. Exact artifact: `orcarouter/Qwen3.8-27B-Uncensored-FP8`, revision `9228df5c6c9c509e1019f83b4e085cf643118bac`. While huihui-ai raced out its "crude proof of concept" abliteration, a second edit quietly became the most-liked Qwen3.8 uncensored on Hugging Face: orcarouter's block-FP8 build, 553 likes in four days, 45,465 downloads, served gated behind an API at $0.40 per million input tokens. Same Qwen3.8-27B base — 27,781,427,952 parameters, 64 hybrid layers, 27-layer vision encoder, 262,144-token context — but a bolder promise: the card calls it lossless, capabilities preserved while refusal behavior is removed, shipped already quantized to block-FP8 with the vision tower at full precision. "Aggressive" here means fully unlocked per the OrcaRouter listing: direct, complete responses across a wide prompt range, occasionally appending brief informational disclaimers inherited from base training — which it states are not refusals. That is a publisher claim; no measured refusal-rate table is published, so this artifact does not carry a ZERO REFUSALS badge. Third-party numbers (Artificial Analysis, evaluated 14 August 2026): AA Coding 68.1 (#30 of 133), AA Intelligence 52.0 (#29 of 135), GPQA Diamond 90.5, Humanity's Last Exam 33.9, Long-Context Recall 77.3, SciCode 44.7, terminalbench v2.1 79.8, tau_banking 48.0. What is missing: a public refusal-reduction measurement for this exact edit and a capability-retention A/B against the unedited base. The edit and serving build come from [orcarouter](https://huggingface.co/orcarouter), the company behind the [OrcaRouter gateway](https://www.orcarouter.ai/) — 197 models, 16 providers, zero token markup. The same crew ships the other "obsidian" uncensored builds (Qwen3.6 35B A3B Aggressive, Gemma4 26B A4B Balanced). ## How to run it The API is OpenAI-compatible and gated: ```python from openai import OpenAI client = OpenAI(base_url="https://api.orcarouter.ai/v1", api_key="ORCAROUTER_API_KEY") resp = client.chat.completions.create(model="obsidian/Qwen3.8-27B", messages=[{"role":"user","content":"Build it."}]) ``` $0.40 per million input tokens, $4.21 per million output, 262K context, image and video input. For self-hosting, the exact pinned repository serves in vLLM on one H200; the ABLITERATED.cloud prepared profile estimates $5.45/hour managed. ## The idea, in plain words **Why FP8 is the boring hero of this story** — Neural networks normally store every weight as a 16-bit number. Block-FP8 packs them into 8 bits with per-block scaling: half the memory, near-identical inference quality, which turns a 55 GB download into a model you can serve on one GPU. Quantization is why "open weights" and "usable API" are different things. Primary sources: - [Exact orcarouter model card](https://huggingface.co/orcarouter/Qwen3.8-27B-Uncensored-FP8) - [Pinned artifact](https://huggingface.co/orcarouter/Qwen3.8-27B-Uncensored-FP8/tree/9228df5c6c9c509e1019f83b4e085cf643118bac) - [OrcaRouter model page — pricing and benchmarks](https://www.orcarouter.ai/models/obsidian/Qwen3.8-27B) - [OrcaRouter gateway](https://www.orcarouter.ai/) and [orcarouter profile](https://huggingface.co/orcarouter) - [Official upstream Qwen card](https://huggingface.co/Qwen/Qwen3.8-27B) - [Artificial Analysis benchmark source](https://artificialanalysis.ai/) "Lossless" and "fully unlocked" are the publisher's claims; the third-party benchmarks are the only independent measurement, and no refusal-rate measurement exists. Unlock comes with the usual trade — you verify what comes back. ## Run this model on your terms Want this model running for you, on a private cloud GPU or your own machine? [Request access on Signal](https://signal.me/#p/+13103408213) or [see how it works](https://abliterated.cloud/#how). > Model research, dated at publication. Model licenses, publisher benchmarks and hosting estimates are specific to each article, not a live availability or price list. Reported zero-refusal results are test-specific, not a universal guarantee.