Qwen 3.8 27B Uncensored | Salient Engineering

Qwen 3.8 27B Uncensored

A self-hosted Qwen 3.8 27B vision-language model in your AWS account. This build has refusals removed and is meant for research and controlled evaluation. Add your own safety layer before any end-user deployment.

Runs only in your account
262K context, thinking mode, and tool calling
Image understanding, including text in images
Apache 2.0 weights, billed by the hour

Free Trial with AWS

Quick start

Check your GPU quota before you subscribe

GPU instances need the EC2 quota Running On-Demand G and VT instances (code L-DB2E81BA) in your Region. It must be at least the vCPU count of the instance (g5.xlarge = 4 vCPU; g5.2xlarge = 8 vCPU; g6.xlarge = 4 vCPU; g6.2xlarge = 8 vCPU; g6e.xlarge = 4 vCPU; g6e.2xlarge = 8 vCPU). New AWS accounts often start at 0, and the launch then fails. Check or request the quota.

Subscribe on AWS Marketplace. The free trial covers the software fee; EC2 charges still apply.
Launch on g6.xlarge. In the security group, allow TCP 443 from the addresses that will call the API, and TCP 22 only if you need SSH.
The current Marketplace AMI keeps weights on disk, so the model is usually ready in about 6 minutes. If you launched an older image that still pulls weights, first boot can take much longer. Leave the instance running and retry /api/tags until the model is listed. In the EC2 console, open the instance, then Actions, Monitor and troubleshoot, Get system log. When the model is ready the log shows the endpoint, the API key, and example requests.
```
curl -k https://PUBLIC_IP/healthz
```
List models with the API key. The certificate is self-signed, so use -k or pin the fingerprint from the system log.
```
curl -k https://PUBLIC_IP/api/tags \
  -H "Authorization: Bearer API_KEY"
```

Generate a completion:

curl -k https://PUBLIC_IP/api/generate \
  -H "Authorization: Bearer API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"orcarouter/Qwen3.8-27B-Uncensored","prompt":"Hello","stream":false}'

Or call the OpenAI-compatible chat endpoint:

curl -k https://PUBLIC_IP/v1/chat/completions \
  -H "Authorization: Bearer API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"orcarouter/Qwen3.8-27B-Uncensored","messages":[{"role":"user","content":"Hello"}]}'

Ollama is bound to localhost. Do not open 11434. Older images that still serve plain HTTP on 11434 need that port limited to your CIDR, or an SSH tunnel (ssh -i key.pem -L 11434:127.0.0.1:11434 ubuntu@PUBLIC_IP).

Instance types on this listing: g5.xlarge, g5.2xlarge, g6.xlarge, g6.2xlarge, g6e.xlarge, g6e.2xlarge. The model name in requests is orcarouter/Qwen3.8-27B-Uncensored.