The same weights, a different promise
Both edits start from the same Qwen3.8-27B: dense, 64 hybrid layers, 27.8 billion parameters, a 27-layer vision encoder and a 262,144-token native context. Where huihui-ai labels its work a crude proof of concept, the orcarouter card calls this one lossless: capabilities preserved, refusal behavior removed, and the whole thing shipped already quantized to block-FP8 — which is why it is the only one of the two you can actually get served today, through OrcaRouter's gated API under the provider label "obsidian".
Being FP8 from the start matters. Huihui's repository is BF16, about 55.6 GB; block-FP8 halves the footprint for serving while keeping the vision tower at full precision, per the card. That is the difference between a weights drop you download and a model you can put behind an endpoint the same day.
What "Aggressive" means here
The OrcaRouter listing describes the aggressive variant as fully unlocked: direct, complete responses across a wide prompt range, with the caveat that the model may occasionally append brief informational disclaimers inherited from base training — which it states are not refusals, since the requested content is still generated in full. That is a publisher claim, not a measured refusal rate. There is no published 0/100-style table for this artifact, which is exactly why this site does not give it a ZERO REFUSALS badge.
The numbers, third-party
OrcaRouter publishes Artificial Analysis results: 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 — evaluated 14 August 2026. These are third-party measurements of the served build, a different class of evidence than a self-reported card. What is still missing: any public measurement of refusal reduction for this exact edit, and any capability-retention A/B against the unedited base.
Who made it
The base is Qwen's. The uncensored edit and the FP8 serving build come from orcarouter, which also runs the OrcaRouter gateway — "one AI gateway: adaptive LLM routing and governance", 197 models across 16 providers, zero token markup. The same crew ships the other "obsidian" uncensored builds (Qwen3.6 35B A3B Aggressive, Gemma4 26B A4B Balanced). They are a serving company first, which explains the shape of the release: API first, gated, benchmarked, priced per token.
How to run it
The API is OpenAI-compatible and gated — sign in, request access, then:
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 directly in vLLM on one H200; our prepared profile estimates $5.45/hour managed, scale-to-zero included — request access and we host it.
The idea, in plain words
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 quality for inference — which turns a 55 GB download into a model you can serve on one GPU at reasonable cost. Quantization is why "open weights" and "usable API" are different things, and why this edit shipped as a service while the BF16 one stayed a download.
How we treat it
Our prepared profile pins revision 9228df5c6c9c509e1019f83b4e085cf643118bac with a conservative 131,072-token context on one H200 at an approximate managed price estimate of $5.45/hour. "Lossless" and "fully unlocked" are the publisher's claims; the third-party benchmarks above are the only independent measurement, and no refusal-rate measurement exists. Unlock comes with the usual trade — you verify what comes back.