Use Inference APIs with opencode

Settings

Which model

A large model for agent work: zai-org/GLM-5.3, deepseek-ai/DeepSeek-V4-Pro or moonshotai/Kimi-K2.7-Code. DeepSeek and GLM here run in the United States, so opencode's "models hosted in China" switch is not needed. If you came here from OpenCode Go's 403 RegionError, that error is explained here. Prices and measured speed: chat models.

Configuration

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "inferenceapis": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Inference APIs",
      "options": {
        "baseURL": "https://api.inferenceapis.com/v1",
        "apiKey": "{env:INFERENCE_API_KEY}"
      },
      "models": {
        "zai-org/GLM-5.3": { "name": "GLM 5.3", "limit": { "context": 1000000, "output": 32768 } },
        "deepseek-ai/DeepSeek-V4-Pro": { "name": "DeepSeek V4 Pro", "limit": { "context": 1000000, "output": 32768 } },
        "deepseek-ai/DeepSeek-V4-Flash": { "name": "DeepSeek V4 Flash", "limit": { "context": 1000000, "output": 32768 } },
        "moonshotai/Kimi-K2.7-Code": { "name": "Kimi K2.7 Code", "limit": { "context": 262144, "output": 32768 } }
      }
    }
  }
}

export INFERENCE_API_KEY="your-key"
opencode            # then /models and pick inferenceapis/zai-org/GLM-5.3

Verify

Run /models inside opencode; the four models appear under "Inference APIs". Send "list the files here" to confirm a tool call works.

How this guide was checked

Run end to end opencode 1.18.31: with exactly the opencode.json below, opencode run read a Python file, fixed a bug in it, wrote a test file and ran the tests, all passing, with DeepSeek V4 Flash (12 seconds), DeepSeek V4 Pro (17 seconds), GLM 5.3 (53 seconds) and Kimi K2.7 Code (100 seconds), on 2026-09-17. The DeepSeek V4 Flash session cost under one cent.

Gotchas

• The model id in opencode is the provider id plus our model id: inferenceapis/zai-org/GLM-5.3.
• Instead of the environment variable you can run /connect and paste the key.
• Set limit.context to the model's real window so opencode compacts the conversation at the right time.
• opencode resends the conversation each turn. The repeated part is served from cache and billed at the cached-input rate, about 80 to 98 percent below normal input, with nothing to configure.

If something fails

401 — key missing or wrong · 402 insufficient_balance — add credits on Billing · 404 model_not_found — check the id against the model list (aliases such as gpt-oss-120b work too) · 503 model_unavailable — the model is being enabled. Full details in the API docs.