Get an API key
An admin adds you to the infer_users group of your shimme.rs ID. Then sign in and create a key on the overview page — it's shown once. Store it in a private file without leaving it in your shell history:
mkdir -p ~/.config/opencode (umask 077; read -rs KEY && printf '%s' "$KEY" > ~/.config/opencode/infer-key)
https://infer.shimme.rs/v1qwen3.6-35b-a3b-uncensoredOpenCode
Add the provider to ~/.config/opencode/opencode.json, merging the providers block if the file exists. OpenCode reloads its config automatically.
{
"$schema": "https://opencode.ai/config.json",
"providers": {
"shimmers": {
"name": "shimme.rs infer",
"package": "@opencode/ai/providers/openai-compatible",
"settings": {
"baseURL": "https://infer.shimme.rs/v1",
"apiKey": "{file:~/.config/opencode/infer-key}",
"headerTimeout": 1860000,
"cache": { "warming": false }
},
"models": {
"qwen3.6-35b-a3b-uncensored": {
"name": "Qwen3.6 35B-A3B Uncensored (shimme.rs GPU)",
"capabilities": {
"tools": true,
"input": ["text"],
"output": ["text"],
"reasoning": { "supported": true }
},
"compatibility": { "reasoningField": "reasoning_content" },
"limit": { "context": 65536, "output": 4096 },
"variants": [
{ "id": "instant", "body": { "chat_template_kwargs": { "enable_thinking": false } } },
{ "id": "thinking", "body": { "chat_template_kwargs": { "enable_thinking": true } } }
]
}
}
}
}
}Run /models and pick Qwen3.6 35B-A3B Uncensored (shimme.rs GPU), or try it from the terminal:
opencode run --model shimmers/qwen3.6-35b-a3b-uncensored "hello"
- Append
#instantto the model for quick answers without thinking. - Keep
cache.warmingoff — keep-alive pings would hold the GPU awake on your allowance. - OpenCode runs tools on your machine; only the conversation reaches the GPU.
- Keep your usual model as default and switch per session to save minutes.
Other tools
Anything that accepts an OpenAI base URL works. Many tools read these variables:
export OPENAI_BASE_URL=https://infer.shimme.rs/v1 export OPENAI_API_KEY="$(cat ~/.config/opencode/infer-key)"
curl
curl https://infer.shimme.rs/v1/chat/completions \
-H "Authorization: Bearer $(cat ~/.config/opencode/infer-key)" \
-H 'Content-Type: application/json' \
-d '{"model":"qwen3.6-35b-a3b-uncensored","stream":true,"messages":[{"role":"user","content":"Hello"}]}'Python
import os
from openai import OpenAI
key = open(os.path.expanduser("~/.config/opencode/infer-key")).read().strip()
client = OpenAI(base_url="https://infer.shimme.rs/v1", api_key=key, timeout=1900, max_retries=0)
stream = client.chat.completions.create(
model="qwen3.6-35b-a3b-uncensored", stream=True,
messages=[{"role": "user", "content": "Hello"}],
)
for chunk in stream:
if chunk.choices:
print(chunk.choices[0].delta.content or "", end="", flush=True)API reference
GET /v1/models | Lists the available model. |
POST /v1/chat/completions | Text messages, streaming, JSON tool calls, response_format, stop. |
| Not supported | Images, embeddings, legacy completions, the Responses API. Sampling parameters are fixed to the model's recommended values. |
Prefer streaming. Streaming requests get headers immediately plus keep-alive comments while the GPU wakes. Non-streaming requests wait silently, so give clients a 30-minute timeout.
Limits & privacy
Scale to zero
One shared GPU wakes on the first request and sleeps after 2 idle minutes. A cold start loads 18.7 GB of weights and takes a few minutes.
Sessions
A GPU session lasts up to 30 minutes; a request still running then is cut off and the next starts a fresh GPU.
Model limits
64k-token context, up to 4,096 output tokens (1,024 for thinking).
Allowance
Default 60 GPU-minutes per month, a shared platform cap and up to 4 queued requests per person. Resets on the 1st (UTC).
Troubleshooting
authentication_error | Key missing, mistyped or revoked. Check the key file has no extra characters. |
| “API access needs renewal” | Sign in here once; membership is re-confirmed every 30 days. |
insufficient_quota | Your or the platform's monthly allowance is used up. |
gpu_unavailable | No GPU within the price limit, a startup failure or cleanup in progress. Wait a minute and resend — clients deliberately don't retry automatically. |
gpu_paused | An admin paused the GPU. |
rate_limit_error | Queue full; retry shortly. |
| First reply is slow | Normal on a cold start; later replies are fast while the GPU is warm. |