Limits & FAQ
The answer says “queued”.
The repository is new to the service. Its files, config and headers are being read from the Hub; ask again after about a minute.
A gated model returns little detail.
Gated configs are not readable anonymously. modelread falls back to the GGUF headers of public quants of the same model when it can, and says so in evidence.
My measured speed is outside the range.
Please report it (see Support) or submit the run with modelfit bench --submit; measurements recalibrate that device class.
Known limits
- MoE offload speeds carry ±50%.
- Quant quality varies ±50% between uploaders at the same label.
- Not modeled: NPU runtimes beyond listing them, speculative-decoding gains, phone thermal throttling, per-GPU fragmentation.
- vLLM fit uses llama.cpp-style memory accounting; the launch line caps the context so it starts.
Rate limits
About 20 requests per minute per client. MCP tool calls count like API calls.
Privacy
No accounts, no cookies of its own. See Privacy.