Tool calling with abliterated models

Tool calling with abliterated models uses the standard formats, OpenAI's tools and tool_calls or Anthropic's tool_use blocks. Kimi K3.1 Pro, DeepSeek V4.1 Flash 1017, GLM-5.3 and GLM-5.3 Flash at Silk Compute all support tool calling with a 1,048,576-token context.

Updated November 18, 2026By unrestrictedaiagent

The two formats

PartChat CompletionsMessages
EndpointPOST /chat/completionsPOST /v1/messages
Define a tooltools[].function.parameterstools[].input_schema
Model asksmessage.tool_calls[]content[] type "tool_use"
You answerrole "tool" with tool_call_id"tool_result" with tool_use_id

OpenAI function calling and Anthropic tool use define each format.

One request in each

Chat Completions
curl https://api.silkcompute.com/chat/completions \
  -H "Authorization: Bearer $SILK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4.1-flash-1017-uncensored",
    "messages": [{"role": "user", "content": "Is the disk on my server almost full?"}],
    "tools": [{"type": "function", "function": {
      "name": "run_command",
      "description": "Run a shell command on my own server",
      "parameters": {"type": "object",
        "properties": {"cmd": {"type": "string"}}, "required": ["cmd"]}}}]
  }'
Messages
curl https://api.silkcompute.com/v1/messages \
  -H "x-api-key: $SILK_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "deepseek-v4.1-flash-1017-uncensored",
    "max_tokens": 4096,
    "messages": [{"role": "user", "content": "Is the disk on my server almost full?"}],
    "tools": [{"name": "run_command",
      "description": "Run a shell command on my own server",
      "input_schema": {"type": "object",
        "properties": {"cmd": {"type": "string"}}, "required": ["cmd"]}}]
  }'

Who sends which

Checks

CheckWhy
Output capOne reply caps at 131,072 tokens on DeepSeek V4.1 Flash and GLM-5.3, and 943,718 on Kimi K3.1 Pro.
Default APIThe OpenAI Agents SDK defaults to Responses, so use its Chat Completions model class.
Responses stateThe Responses endpoint here is stateless, so send the full history each turn.