Kimi K3

Kimi K3 uses HopBase's OpenAI Chat Completions compatible protocol. Use https://api.hop-base.com/v1 and set the model to kimi-k3.

Group selection and billing

• When creating or editing an API key under API Keys in the console, add the Kimi K3 Cloud group to the key. That key can then call kimi-k3.
• Your effective rate is shown in the signed-in model catalog.
• Keys without this group get a clear 404 when calling kimi-k3; add the group to the key in the console.

Client fields

Client support

Kimi K3 serves POST /v1/chat/completions only. Both /v1/responses and /v1/messages return 404 for this model, so clients that only speak the OpenAI Responses protocol or the Anthropic Messages protocol need a protocol translation layer.

Request parameters

Measured parameter by parameter through the production gateway on 2026-10-10. The gateway rejects empty messages and removes the parameters marked "accepted, ignored" below; every other limit comes from the model service.

Use Kimi K3 in Claude Code with CC Switch

Use CC Switch 3.17.0 or later (the latest release is recommended). Kimi K3 uses OpenAI Chat Completions, so CC Switch must translate requests through its local router into the Anthropic Messages protocol used by Claude Code.

Use Kimi K3 in Codex CLI with CC Switch

This also requires CC Switch 3.17.0 or later. CC Switch keeps Codex talking Responses to the local router (default 127.0.0.1:15721); the router rewrites each request into Chat Completions for HopBase and converts the response back into Responses form.

Thinking output, prefix caching, and tool calls

Measured by HopBase through the production gateway in October 2026:

• Kimi K3 currently returns no thinking trace. Responses carry no reasoning_content field and usage.completion_tokens_details.reasoning_tokens is always 0, so every output token is answer text. Sending thinking or reasoning_effort does not fail, but has no effect. Size max_tokens for the answer alone (up to 32,768); nothing needs to be reserved for thinking.
• Long context works and retrieves accurately. A single input of about 470,000 tokens returned normally, and a fact buried in it was retrieved exactly.
• Prefix caching hits the whole prefix. Repeating the same prefix makes usage.prompt_tokens_details.cached_tokens equal to the matched prefix length (a 2,093-token system prompt hits all 2,093 on the second call), billed at the cached rate. Keep the system prompt, tool definitions, and other fixed content at the front of the message list.
• Tool calling works. tools, tool_choice (including required), parallel_tool_calls, JSON Schemas using $ref and $defs, and multi-turn tool messages all behave normally in both streaming and non-streaming mode.
• Latency reference. In a 10-minute sustained load test the median time to first token was about 1.7 seconds, and a 9,000-token non-streaming answer returned in about 78 seconds. Keep client retries and a read timeout of at least 300 seconds.

OpenAI SDK / Python

from openai import OpenAI

client = OpenAI(
    base_url="https://api.hop-base.com/v1",
    api_key="sk-your-key",
)

resp = client.chat.completions.create(
    model="kimi-k3",
    max_tokens=4096,
    messages=[{"role": "user", "content": "Explain where Kimi K3 works best"}],
)
print(resp.choices[0].message.content)
print(resp.usage)

curl

curl https://api.hop-base.com/v1/chat/completions \
  -H "Authorization: Bearer sk-your-key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kimi-k3",
    "max_tokens": 4096,
    "messages": [{"role": "user", "content": "Hello, introduce yourself"}]
  }'

FAQ

• /v1/responses or /v1/messages returns 404. This model only has /v1/chat/completions. Route Codex CLI and Claude Code through CC Switch as described above; do not put the HopBase address directly into ~/.codex/config.toml or ANTHROPIC_BASE_URL.
• Codex fails to start with wire_api = "chat" is no longer supported. Newer Codex no longer supports direct Chat Completions. Remove the hand-written provider config and let CC Switch take over.
• Calling kimi-k3 returns 404 model not found. The key does not include the Kimi K3 Cloud group; add the group to the key or create a new key.
• The Codex /model menu does not list kimi-k3. Restart Codex after saving the model mapping so the menu refreshes.
• A 400 that only says to check the request parameters. Look first at a temperature above 1, an n above 1, or an image sent as a public URL instead of base64.
• A 429 saying the concurrency limit is reached. That is your account concurrency limit, 5 by default. Ask us to raise it before running concurrency benchmarks or multi-agent workloads.

Next steps

• OpenAI SDK — the general Chat Completions pattern
• CC Switch — installing and troubleshooting the local router
• Error codes and retries
• Concurrency, timeouts, and billing