[Bug] Plugin Not Detecting Chat Models - Missing nemotron and glm in categorizeModel()

[oliveiraantoniocc]

[Bug] OpenCode Plugin Not Detecting Chat Models - Incorrectly Classifies nemotron and glm as Unknown

Summary

The opencode-lmstudio plugin fails to detect certain chat models (nvidia/nemotron and zai-org/glm) as chat-capable models, resulting in a warning message that only embedding models are available even when valid chat models are loaded in LM Studio.

Current Behavior

When using LM Studio with OpenCode integration through the opencode-lmstudio plugin, users see this warning:

[opencode-lmstudio] Only embedding models found. To use chat models:
{
  steps: [
    "1. Open LM Studio application",
    "2. Download a chat model (e.g., llama-3.2-3b-instruct)",
    "3. Load the model in LM Studio",
    "4. Ensure server is running"
  ]
}

Even when valid chat models are already loaded and working:

• nvidia/nemotron-3-nano
• zai-org/glm-4.6v-flash
• qwen/qwen3-vl-30b (this one works correctly)

Environment

• OpenCode: Latest version (via Bun)
• Plugin: opencode-lmstudio@0.3.0 (installed globally via bun add -g opencode-lmstudio)
• LM Studio: Running locally on http://127.0.0.1:1234/v1
• Available Models:
  • nvidia/nemotron-3-nano (chat model - NOT detected)
  • zai-org/glm-4.6v-flash (chat/vision model - NOT detected)
  • qwen/qwen3-vl-30b (chat/vision model - correctly detected)
  • text-embedding-nomic-embed-text-v1.5 (embedding model - correctly detected)

Root Cause Analysis

I analyzed the plugin source code and found the issue:

File: ~/.bun/install/global/node_modules/opencode-lmstudio/src/utils/index.ts
Function: categorizeModel() (lines 6-18)

export function categorizeModel(modelId: string): 'chat' | 'embedding' | 'unknown' {
  const lowerId = modelId.toLowerCase()
  if (lowerId.includes('embedding') || lowerId.includes('embed')) {
    return 'embedding'
  }
  if (lowerId.includes('gpt') || lowerId.includes('llama') || 
      lowerId.includes('claude') || lowerId.includes('qwen') ||
      lowerId.includes('mistral') || lowerId.includes('gemma') ||
      lowerId.includes('phi') || lowerId.includes('falcon')) {
    return 'chat'
  }
  return 'unknown'  // ← nemotron and glm fall through to here!
}

The Problem: The function only checks for specific model family names. Models like nvidia/nemotron-3-nano and zai-org/glm-4.6v-flash don't contain any of these keywords, so they're classified as 'unknown'.

File: src/plugin/enhance-config.ts (lines 70-108)

The plugin counts models by category when discovering them, but only counts newly discovered models. When models are pre-configured in opencode.json, they're in existingModels and skipped by the counting logic.

Additionally, on lines 122-131, the warning is triggered when chatModelsCount === 0 && embeddingModelsCount > 0, which happens when:

1. All chat models are pre-configured (not counted as "discovered")
2. Only the embedding model is newly discovered
3. Result: Warning shown even though chat models exist

Steps to Reproduce

1. Install LM Studio and load these models:

   • nvidia/nemotron-3-nano
   • zai-org/glm-4.6v-flash
   • qwen/qwen3-vl-30b

2. Install opencode-lmstudio plugin:

   bun add -g opencode-lmstudio@latest

3. Create opencode.json:

   {
     "$schema": "https://opencode.ai/config.json",
     "plugin": ["opencode-lmstudio@latest"],
     "provider": {
       "lmstudio": {
         "npm": "@ai-sdk/openai-compatible",
         "name": "LM Studio",
         "options": {
           "baseURL": "http://127.0.0.1:1234/v1"
         }
       }
     }
   }

4. Run OpenCode:

   opencode web

5. Observe: Warning appears about only finding embedding models

6. Verify models work:

   # Test with curl - works perfectly!
   curl -X POST http://localhost:1234/v1/chat/completions \
     -H "Content-Type: application/json" \
     -d '{
       "model": "nvidia/nemotron-3-nano",
       "messages": [{"role": "user", "content": "Hello"}]
     }'
   # ✅ Returns valid response

Proposed Solution

Fix 1: Add Missing Model Families to categorizeModel()

File: src/utils/index.ts, lines 11-14

// Current:
if (lowerId.includes('gpt') || lowerId.includes('llama') || 
    lowerId.includes('claude') || lowerId.includes('qwen') ||
    lowerId.includes('mistral') || lowerId.includes('gemma') ||
    lowerId.includes('phi') || lowerId.includes('falcon')) {
  return 'chat'
}

// Proposed fix - add nemotron and glm:
if (lowerId.includes('gpt') || lowerId.includes('llama') || 
    lowerId.includes('claude') || lowerId.includes('qwen') ||
    lowerId.includes('mistral') || lowerId.includes('gemma') ||
    lowerId.includes('phi') || lowerId.includes('falcon') ||
    lowerId.includes('nemotron') || lowerId.includes('glm')) {  // ← ADD THESE
  return 'chat'
}

Fix 2: Count Existing Configured Models

File: src/plugin/enhance-config.ts, after line 68

// Add BEFORE the model discovery loop:
// First, count existing configured models
for (const existingKey of Object.keys(existingModels)) {
  const existingModel = existingModels[existingKey]
  const modelId = existingModel.id || existingKey
  const modelType = categorizeModel(modelId)
  if (modelType === 'chat') {
    chatModelsCount++
  } else if (modelType === 'embedding') {
    embeddingModelsCount++
  }
}

This ensures pre-configured models are counted, preventing false warnings.

Verification

After applying Fix 1 manually to test:

# Test script showing the fix works:
cat > /tmp/test-categorize.ts << 'EOF'
function categorizeModel(modelId: string): 'chat' | 'embedding' | 'unknown' {
  const lowerId = modelId.toLowerCase()
  if (lowerId.includes('embedding') || lowerId.includes('embed')) {
    return 'embedding'
  }
  if (lowerId.includes('gpt') || lowerId.includes('llama') || 
      lowerId.includes('claude') || lowerId.includes('qwen') ||
      lowerId.includes('mistral') || lowerId.includes('gemma') ||
      lowerId.includes('phi') || lowerId.includes('falcon') ||
      lowerId.includes('nemotron') || lowerId.includes('glm')) {
    return 'chat'
  }
  return 'unknown'
}

console.log('nvidia/nemotron-3-nano ->', categorizeModel('nvidia/nemotron-3-nano'))
console.log('zai-org/glm-4.6v-flash ->', categorizeModel('zai-org/glm-4.6v-flash'))
console.log('qwen/qwen3-vl-30b ->', categorizeModel('qwen/qwen3-vl-30b'))
EOF

bun run /tmp/test-categorize.ts

Output:

nvidia/nemotron-3-nano -> chat     ✅
zai-org/glm-4.6v-flash -> chat     ✅
qwen/qwen3-vl-30b -> chat          ✅

Impact

Who This Affects:

• Users of LM Studio with OpenCode
• Anyone using nvidia/nemotron or zhipu/glm model families
• Anyone who pre-configures models in opencode.json

Current Workaround:
The models actually work fine! The warning is incorrect. Users can:

1. Ignore the warning
2. Manually specify models in config with explicit capabilities:

   "models": {
     "nvidia/nemotron-3-nano": {
       "id": "nvidia/nemotron-3-nano",
       "name": "Nemotron 3 Nano",
       "tool_call": true,
       "temperature": true
     }
   }

After Fix:

• No false warnings
• Automatic detection of more model families
• Better user experience

Additional Context

Popular Models That Would Benefit:

• NVIDIA Nemotron series (3B-70B+ parameters)
• Zhipu AI GLM series (ChatGLM, GLM-4, etc.)
• Future model families not in the hardcoded list

Suggested Long-Term Improvement:

Instead of hardcoding model names, consider using the LM Studio API's model metadata or a more flexible pattern matching approach that:

1. Checks if the model supports /v1/chat/completions endpoint
2. Looks for chat-related capabilities in the model metadata
3. Falls back to embedding if it only supports /v1/embeddings

This would make the plugin more maintainable and future-proof.

Related

• Main plugin repository: https://github.com/agustif/opencode-lmstudio
• Model families affected: NVIDIA Nemotron, Zhipu GLM
• LM Studio: https://lmstudio.ai/

---

Labels: bug, opencode-lmstudio
Priority: Low (cosmetic warning, functionality works)

Let me know if you need any additional information or testing!