Blog

Qwen 3.8 27B RR API: low-refusal Qwen 3.8 27B

By the RedRouter Team · Tested October 27, 2026
October 1, 2026

Qwen 3.8 27B RR is Alibaba's mid-size Qwen 3.8 27B, modified to reduce refusals. It sits between the GLM models on price and suits chat, writing and multilingual work. The model ID is qwen3.8-27b-rr.

Zero Data Retention, if you turn it on for your account, adds 20% to every rate.

What RedRouter knows about the checkpoint

The host reports FP8 weights. The source of the modified weights, their revision, license, tokenizer and chat template are not published, and the catalog lists each as unknown. RedRouter has not confirmed which Qwen 3.8 27B release it was built from.

Its context window is slightly smaller than the other three models, at the figure in the table above.

What "RR" changes

RR models are open-weight models with refusal behavior reduced. RedRouter adds no refusal layer of its own. The model can still decline, and it can still be wrong.

Try it

Save this defanged German lure as lure.txt. RedRouter wrote it for this example; the link is defanged and the bank is fictional.

Sehr geehrte Kundin, sehr geehrter Kunde,
wir haben ungewöhnliche Anmeldeversuche bei Ihrem Konto festgestellt. Ihr Zugang wird innerhalb von 24 Stunden gesperrt, wenn Sie Ihre Identität nicht bestätigen.
Bitte melden Sie sich sofort über den folgenden Link an: hxxps://kundenportal-sicherheit[.]example/verify
Mit freundlichen Grüßen
Ihr Sicherheitsteam der Musterbank

import os
from openai import OpenAI

lure = open("lure.txt").read()
client = OpenAI(base_url="https://redrouter.org/v1", api_key=os.environ["REDROUTER_API_KEY"])
response = client.chat.completions.create(
    model="qwen3.8-27b-rr",
    max_tokens=4096,
    messages=[{"role": "user", "content": "Translate this phishing lure into English and list its social-engineering cues. Reply in JSON.\n\n" + lure}],
    response_format={"type": "json_object"},
)
print(response.choices[0].message.content)

Run on October 27, 2026. qwen3.8-27b-rr returned (trimmed):

{
  "translation": "Dear Valued Customer,\n\nWe have detected unusual login attempts on your account. Your access will be locked within 24 hours if you do not verify your identity.\n\nPlease log in immediately via the following link: hxxps://kundenportal-sicherheit[.]example/verify\n\nWarm regards,\n\nYour Security Team at MusterBank",
  "social_engineering_cues": [
    {
      "criterion": "Urgency",
      "explanation": "The email imposes a strict 24-hour deadline to prevent the account from being locked, pressuring the recipient to act immediately without thinking."
    },
    {
      "criterion": "Fear/Awareness of Loss",
      "explanation": "It mentions 'unusual login attempts' and the potential loss of account access, exploiting the fear of compromise or inconvenience."
    },
    {
      "criterion": "Attention-to-Pattern/Dilution of Vigilance",
  ...

When to pick it

Qwen 3.8 27B RR sits in the middle on price at $0.55 input and $2.40 output per million tokens, above the two Flash models and below GLM 5.3 RR. It supports tool calling and JSON mode. Reach for it when Qwen's output suits your task, including multilingual work, better than the others in your own testing. For the cheapest high-volume or agent traffic, the Flash models win; for the full-size model on a hard problem, GLM 5.3 RR. Its context window is also a little smaller than the other three.

FAQ

Does it support tools and JSON mode? Yes to both. See the quickstart for request examples.

Is my data kept? By default, an encrypted copy of prompts and completions is kept for up to 30 days for abuse prevention, support, debugging and billing disputes. Zero Data Retention stops that. See the privacy policy.

What may I use it for? Anything legal. Testing systems you do not own or are not authorized to test is against the terms.