ollama-cloud: glm-5.3-flash dropped from model cache due to unmapped family 'glm5_next'

Description

glm-5.3-flash:cloud (released on Ollama 2026-08-26) is silently dropped from opencode's ollama-cloud model cache and never appears in opencode models ollama-cloud or the TUI model picker, even though it is registered in ~/.ollama/config.json and reachable on Ollama's API.

Root cause: opencode's ollama-cloud cache builder normalizes each model's family field from Ollama's /api/show response to a canonical family name. The new GLM-5.3-Flash architecture reports family: "glm5_next", which has no mapping in opencode's family normalization table, so the model is silently discarded during cache construction.

• Existing GLM models map glm5.2 → glm and glm5.1 → glm.
• glm5_next is a new architecture (Z.ai redesigned GLM-5.3 with linear + sparse attention, 320B total / 18B active, first natively multimodal GLM) and needs its own entry.
• The model supports vision (capabilities: completion, thinking, tools, vision), so attachment: true should be set — currently glm-5.2 is attachment: false.

Plugins

None.

OpenCode version

1.18.23

Steps to reproduce

1. Run ollama launch opencode --model glm-5.3-flash:cloud. This writes ~/.ollama/config.json:

   {
     "integrations": {
       "opencode": {
         "models": ["glm-5.3-flash:cloud", "glm-5.2:cloud"]
       }
     }
   }
   

   Confirmed working via ollama's own launcher.
2. Run opencode models ollama-cloud. Expected: glm-5.3-flash listed. Actual: absent (only glm-5.2 and glm-5.1 appear).
3. Run opencode run --model ollama-cloud/glm-5.3-flash:cloud "say hi". Actual: fails with:

   Error: {
     "name": "UnknownError",
     "data": {
       "message": "Unexpected server error. Check server logs for details.",
       "ref": "err_60ae5eb7"
     }
   }
   

4. Verify the model is reachable on Ollama's API (works):

   curl -s "https://ollama.com/v1/models/glm-5.3-flash:cloud"
   => {"id":"glm-5.3-flash:cloud","object":"model","created":1787752800,"owned_by":"ollama"}
   
   curl -s "https://ollama.com/api/show" -d '{"name":"glm-5.3-flash:cloud"}'
   => {
     "capabilities": ["completion", "thinking", "tools", "vision"],
     "details": {
       "parent_model": "glm-5.3-flash:cloud",
       "family": "glm5_next",
       "parameter_size": "321323031390",
       "quantization_level": "FP8"
     },
     "model_info": {
       "general.architecture": "glm5_next",
       "general.parameter_count": 321323031390,
       "glm5_next.context_length": 1048576,
       "glm5_next.embedding_length": 4096
     },
     "modified_at": "2026-08-26T07:00:00-07:00"
   }
   

5. Inspect ~/.cache/opencode/models.json — the ollama-cloud provider has families:

   ['deepseek-flash', 'deepseek-thinking', 'gemma', 'glm', 'gpt-oss', 'kimi-k2', 'kimi-k3', 'minimax', 'minimax-m3', 'mistral-large', 'nemotron', 'qwen']
   

   No glm5_next / glm5.3 present. glm-5.2 (works) reports family: "glm5.2" → normalized to glm; glm-5.3-flash (dropped) reports family: "glm5_next" → no normalization entry.

Screenshot and/or share link

N/A — reproduction is fully CLI-driven; outputs included in Steps to reproduce.

Operating System

macOS (darwin)

Terminal

iTerm2

Suggested fix

Add glm5_next → glm (or a new glm5-next family) to the ollama family normalization map used by the ollama-cloud cache builder. Set attachment: true for this model since GLM-5.3-Flash supports vision (the first natively multimodal GLM).