Thinking, fast and slow.
AI that decides in a blink and reasons in depth.
Save it now: it is shown only once.
curl -s https://app.vllm-sr.ai/v1/systemone \
-H "Authorization: Bearer $VLLM_SR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model": "vllm-sr/d3-ultra", "state": "Our app has returned 500 errors since the 09:00 deploy.",
"questions": {"route": {"type": "choice", "instructions": "Which team should handle this ticket?",
"criteria": {"billing": "Payments and refunds", "engineering": "Outages and bugs"}}}}'curl -s https://app.vllm-sr.ai/v1/chat/completions \
-H "Authorization: Bearer $VLLM_SR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model": "vllm-sr/auto", "messages": [{"role": "user", "content": "Hello!"}]}'Requests are logged to improve the service, so do not send sensitive data. Privacy
Live
Pick an example. It runs live.
Every answer comes with its probability.
The router picks the model and the effort.
System Two is not reachable right now. System One still answers; try again in a minute.
Pick an example to see the question.
A coding agent with System One at its side, in a sandbox of your own.
DeepSeek Harness, live
Read-only example sessions, and sessions of your own in a sandbox of your own, on your own free API key.
Public demo · Terms · Privacy · Open in a new tab ·
vLLM Semantic Router serves System One and routes System Two.
System Two · every chat request
Any OpenAI SDK, model="vllm-sr/auto"
The model and its reasoning effort, from the signals.
Streaming, tools and the routing headers.
Any model, open or closed
System One · a direct question
POST /v1/systemone: risky? Which team? Done?
d3 and Vela 3.0: text, images and video; spans in text.
A probability for each option, no generated text.
Allow, ask, route or stop, at your threshold.
The agent’s LLM keeps its tokens for the work.
From your agent or your terminal.
DeepSeek Harness
Add the plugin; it gets its own free key. vllm-sr/auto plans and acts, System One guards and checks.
npm install -g @deepseek-ai/dsh pnpm
dsh plugin --profile web add @vllm-sr/dsh@0.6.11
dsh webGet a key, set it once, call either system.
POST /v1/systemonePOST /v1/chat/completionsAuthorization: Bearerexport VLLM_SR_API_KEY=sk-vsr-...
# System One: a typed question, a probability for every option
curl -s https://app.vllm-sr.ai/v1/systemone \
-H "Authorization: Bearer $VLLM_SR_API_KEY" -H "Content-Type: application/json" \
-d '{"model": "vllm-sr/d3-ultra", "state": "Refund my order, it arrived broken.",
"questions": {"team": {"type": "choice", "instructions": "Which team handles this?",
"criteria": {"returns": "Refunds and returns", "tech": "Product faults"}}}}'
# System Two: chat, routed to the right model
curl -s https://app.vllm-sr.ai/v1/chat/completions \
-H "Authorization: Bearer $VLLM_SR_API_KEY" -H "Content-Type: application/json" \
-d '{"model": "vllm-sr/auto", "messages": [{"role": "user", "content": "Plan a 3-day trip to Kyoto."}]}'Open models, open routing, open ROCm, on AMD Instinct.
DeepSeek Harnesssk-vsr-…Free, from one click. Every model, one origin: app.vllm-sr.ai
Open weights on Hugging Face
On the open ROCm software stack
Four decision models. One routed model name.
System One · fast
POST /v1/systemone
| Model | Size | Text / vision | Model card |
|---|---|---|---|
| Decision 3.0decisions on text, images and video | |||
| d3-ultra | 27B | 70.18 / 70.59 | Card |
| d3-flash | 9B | 54.27 / 66.89 | Card |
| Vela 3.0d3’s decisions, plus spans and router signals | |||
| vela3-turbo | 27B | 70.3 / 72.8 est.via its d3‑turbo base | Card |
| vela3-flash | 9B | 54.27 / 66.89via its d3‑flash base | Card |
System Two · slow
vllm-sr/autoPOST /v1/chat/completions
One OpenAI-compatible model name. vLLM Semantic Router sends each request to the model and reasoning effort it needs.
Works with any model, open or closed
GPT
Claude
Gemini
DeepSeek
Qwen
GLM
Llama
Mistral
Gemma
MiniMaxServing now
GLM-5.3-Flashzai/glm-5.3-flashQwen3.8-Flash-Nextqwen/qwen3.8-flash-nextQwen3.8-27Bqwen/qwen3.8-27bStreaming, tools and images; the routing decision rides in every answer’s headers.
System Two docsMission
Every request deserves the right model: not always the biggest, never a guess.
One typed call answers every routing question before the first token.
Thinking is a cost. Spend it on the proof, not on the fun fact.
Open source, open models and the open ROCm stack on AMD Instinct.