Strange output when running opencode

[DarkEye123]

I tried Qwen 3.5 35b A3B today with opencode and it seems something is off

It seems like endpoint isn't fully openai compatible. Atleast it seems it doesn't properly accept configuration from opencode.

I didn't yet receive reasonable answer from the model. Can you help how to set it up properly? I even tried different chat template from unsloth, but didn't help with the issue

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "llama.cpp": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "llama-server (local)",
      "options": {
        "baseURL": "http://127.0.0.1:8080/v1"
      },
      "models": {
        "claude": {
          "name": "local claude model",
          "modalities": { "input": ["image", "text"], "output": ["text"] },
          "limit": {
            "context": 262144,
            "output": 262144
          }
        },
        "q-3.5-35b-a3b": {
          "name": "Q-3.5-35B-A3B",
          "modalities": { "input": ["image", "text"], "output": ["text"] },
          "limit": {
            "context": 262144,
            "output": 81920
          },
          "cost": {
            "input": 0.3125,
            "output": 1.25,
            "cache_read": 0.1563
          },
          "temperature": true,
          "variants": {
            "general": {
              "options": {
                "temperature": 1.0,
                "top_p": 0.95,
                "top_k": 20,
                "min_p": 0.0,
                "presence_penalty": 1.5,
                "repetition_penalty": 1.0,
                "chat_template_kwargs": {
                  "enable_thinking": true
                }
              }
            },
            "coding": {
              "options": {
                "temperature": 0.6,
                "top_p": 0.95,
                "top_k": 20,
                "min_p": 0.0,
                "presence_penalty": 0.0,
                "repetition_penalty": 1.0,
                "chat_template_kwargs": {
                  "enable_thinking": true
                }
              }
            },
            "no-thinking": {
              "options": {
                "temperature": 0.7,
                "top_p": 0.8,
                "top_k": 20,
                "min_p": 0.0,
                "presence_penalty": 1.5,
                "repetition_penalty": 1.0,
                "chat_template_kwargs": {
                  "enable_thinking": false
                }
              }
            },
            "reasoning": {
              "options": {
                "temperature": 1.0,
                "top_p": 1.0,
                "top_k": 40,
                "min_p": 0.0,
                "presence_penalty": 2.0,
                "repetition_penalty": 1.0,
                "chat_template_kwargs": {
                  "enable_thinking": false
                }
              }
            }
          }
        }
      }
    },
    "krasis": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "krasis-server (local)",
      "options": {
        "baseURL": "http://localhost:8012/v1"
      },
      "models": {
        "q-3.5-35b-a3b": {
          "name": "Q-3.5-35B-A3B",
          "modalities": { "input": ["image", "text"], "output": ["text"] },
          "limit": {
            "context": 262144,
            "output": 81920
          },
          "cost": {
            "input": 0.3125,
            "output": 1.25,
            "cache_read": 0.1563
          },
          "temperature": true,
          "variants": {
            "general": {
              "options": {
                "temperature": 1.0,
                "top_p": 0.95,
                "top_k": 20,
                "min_p": 0.0,
                "presence_penalty": 1.5,
                "repetition_penalty": 1.0,
                "chat_template_kwargs": {
                  "enable_thinking": true
                }
              }
            },
            "coding": {
              "options": {
                "temperature": 0.6,
                "top_p": 0.95,
                "top_k": 20,
                "min_p": 0.0,
                "presence_penalty": 0.0,
                "repetition_penalty": 1.0,
                "chat_template_kwargs": {
                  "enable_thinking": true
                }
              }
            },
            "no-thinking": {
              "options": {
                "temperature": 0.7,
                "top_p": 0.8,
                "top_k": 20,
                "min_p": 0.0,
                "presence_penalty": 1.5,
                "repetition_penalty": 1.0,
                "chat_template_kwargs": {
                  "enable_thinking": false
                }
              }
            },
            "reasoning": {
              "options": {
                "temperature": 1.0,
                "top_p": 1.0,
                "top_k": 40,
                "min_p": 0.0,
                "presence_penalty": 2.0,
                "repetition_penalty": 1.0,
                "chat_template_kwargs": {
                  "enable_thinking": false
                }
              }
            }
          }
        }
      }
    }
  },
  "mcp": {
    "context7": {
      "type": "local",
      "command": ["npx", "-y", "@upstash/context7-mcp"]
    }
  }
}

[brontoguana]

I think there are some subtleties here with how Krasis is interacting with the opencode client. I think there may be more than one area that needs some work. I'm currently trying to finish off more robust multi-GPU support and increased decode speed, then I'll take a look at how thinking is handled particularly in opencode and other tool based clients like Aider and try to get them working better with Krasis.

[jwcrawley]

Awesome work btw, @brontoguana . I'm seeing triple my token output and also running a class heavier model. Frankly this is bonkers what you're doing :)

But I think part of this problem might be the string vs array OpenAI API compatibility issue. I just opened up an issue as well, with 2 sample tests commands.