eval_type: inference-check
model: Nanthasit/sakthai-plus-1.5b
timestamp: 2026-07-31T00:29:43Z
results:
  - method: curl POST to api-inference.huggingface.co
    status: dns_unreachable
    detail: "api-inference.huggingface.co does not resolve in sandbox DNS (gaierror -5)"
    http_status: null
    response_time_sec: null
    output: null

  - method: InferenceClient with provider='auto'
    status: no_provider_mapping
    detail: "Model has empty inference_provider_mapping; StopIteration in provider selection"
    http_status: null
    response_time_sec: 0.134
    output: null

  - method: InferenceClient with provider='hf-inference'
    status: model_not_supported
    detail: "BadRequestError: Model not supported by provider hf-inference"
    http_status: 400
    response_time_sec: 0.265
    output: '{"error":"Model not supported by provider hf-inference"}'

  - method: router.huggingface.co/v1/chat/completions
    status: model_not_supported
    detail: "Model not supported by any enabled provider"
    http_status: 400
    response_time_sec: 0.148
    output: '{"error":{"message":"The requested model is not supported by any provider you have enabled.","code":"model_not_supported"}}'

  - method: local transformers inference
    status: oom
    detail: "OOM (exit 137) — sandbox has 7.8GB RAM, 712MB free; 1.5B model requires ~3GB (fp16) or ~6GB (fp32)"
    http_status: null
    response_time_sec: null
    output: null

summary:
  accessible: false
  root_cause: |
    The old inference API (api-inference.huggingface.co) is fully deprecated and has no DNS records.
    The new Inference Providers router rejects the model because no provider has it in their catalog.
    Local inference impossible due to memory constraints (712MB free).
  recommendation: |
    To get this model serving inference, either:
    a) Enable serverless inference for the model on HF Hub (Settings → Inference), which makes
       hf-inference provider load it on-demand.
    b) Deploy a dedicated Inference Endpoint ($$ — not compatible with Zero-Cost First principle).
    c) Convert to GGUF and run via llama.cpp on a machine with ≥4GB RAM.