Dev setup — self-hosted coding model

Connection details

Shared gateway, private repo — don't paste your key anywhere public.

The base URL is permanent. The gateway stays put and forwards to whichever GPU box is running that day. You never need a new address again.

Your key is yours. Each developer has their own, so usage is attributed per person and a key can be revoked without disturbing anyone else.

curl -s https://gpu.revvitysignalsai.com/v1/models \
  -H "Authorization: Bearer YOUR-KEY" | python3 -m json.tool

Which model is which

This is the full catalog of models we have run or may run — not the live set. Which of them are actually served changes from time to time, so always confirm against GET /v1/models before configuring one.

Checkpoint size is what gets downloaded at boot, and it's the main reason a cluster takes 20–60 minutes to come up. Context is what you can actually send.

Not measured means exactly that — we haven't run it long enough to read max_model_len off a live pod, and a number invented here would be worse than an honest gap. Read the real value any time:

curl -s https://gpu.revvitysignalsai.com/v1/models \
  -H "Authorization: Bearer YOUR-KEY" | python3 -m json.tool

opencode

~/.config/opencode/opencode.json — copy-paste, this is the working config. The model entries below are a snapshot of what was live when it was written — the set changes over time, so check GET /v1/models and add, remove or adjust entries to match what is actually served:

{
  "$schema": "https://opencode.ai/config.json",
  "disabled_providers": ["amazon-bedrock", "huggingface"],
  "provider": {
    "gpu-poc": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "GPU POC (self-hosted)",
      "options": {
        "baseURL": "https://gpu.revvitysignalsai.com/v1",
        "apiKey": "YOUR-KEY-HERE"
      },
      "models": {
        "qwen3.8-flash-next": {
          "name": "Qwen3.8 Flash Next (vision)",
          "limit": { "context": 262144, "output": 16384 },
          "tool_call": true,
          "reasoning": true,
          "attachment": true,
          "modalities": { "input": ["text", "image"], "output": ["text"] }
        },
        "deepseek-v41-flash": {
          "name": "DeepSeek V4.1 Flash (vision)",
          "limit": { "context": 1048576, "output": 16384 },
          "tool_call": true,
          "reasoning": true,
          "attachment": true,
          "modalities": { "input": ["text", "image"], "output": ["text"] },
          "options": { "reasoning_effort": "high" }
        },
        "glm-5.3-flash": {
          "name": "GLM-5.3 Flash",
          "limit": { "context": 262144, "output": 16384 },
          "tool_call": true,
          "reasoning": true
        }
      }
    }
  },
  "agent": {
    "default": { "model": "gpu-poc/qwen3.8-flash-next" },
    "plan":    { "model": "gpu-poc/qwen3.8-flash-next" },
    "build":   { "model": "gpu-poc/qwen3.8-flash-next" }
  }
}

qwen3.8-flash-next is set as the default for every agent, so opencode uses it without you picking anything. Switch per-session with /models → GPU POC (self-hosted) — every model you declared appears there. If a model you want is missing, discover what the gateway is serving with GET /v1/models, add it to the config, and restart opencode.

Each limit.context matches what that pod actually serves (262,144 for Qwen3.8; 1,048,576 for DeepSeek V4.1). A client that claims more than the backend serves fails mid-conversation rather than at connect time, which is a miserable way to find out — so keep these equal to max_model_len from GET /v1/models, not higher.

Both attachment and modalities are required for images. attachment: true only permits a file to be attached; modalities.input containing "image" is what makes opencode send it as an image. With attachment alone, opencode degrades the file to a text part and the model replies "I can't view images — this model doesn't support image input" — which is true of what it received, not of the model. Verified 2026-09-14.

Only declare vision on models that have it. On a text-only model the image is dropped and you get a confident answer about nothing.

Config changes need an opencode restart. It reads config at startup.

disabled_providers keeps the picker clean. opencode auto-detects providers from environment credentials, so AWS creds pull in Bedrock and an HF_TOKEN pulls in HuggingFace — both then compete with our models in the list and get selected by accident.

Any other editor

Cursor, Continue, Zed, Cline — wherever it asks for an OpenAI-compatible or custom provider:

Aider

aider --openai-api-base https://gpu.revvitysignalsai.com/v1 \
      --openai-api-key YOUR-KEY-HERE \
      --model openai/qwen3.8-flash-next

Sending an image

Standard OpenAI multimodal format, against a vision model only (qwen3.8-flash-next or deepseek-v41-flash — not glm-5.3-flash, which is text-only):

curl https://gpu.revvitysignalsai.com/v1/chat/completions \
  -H "Authorization: Bearer YOUR-KEY" -H 'Content-Type: application/json' \
  -d '{"model":"qwen3.8-flash-next","max_tokens":500,
       "messages":[{"role":"user","content":[
         {"type":"image_url","image_url":{"url":"data:image/png;base64,..."}},
         {"type":"text","text":"What does this screenshot show?"}]}]}'

image_url takes a data: URI or a public URL. Base64 counts against your context — a full-screen screenshot is often several thousand tokens.