Integration#
Each tool below has its own page. Omit --cache-dir: the model downloads into
the Hugging Face cache, and the engine is reused from the default Edge-LLM
cache.
edgellm serve Qwen/Qwen3.5-4B \
--max-input-len 32768 \
--max-kv-cache-capacity 40960 \
--served-model-name qwen35-local \
--port 8000
Generation clients use --max-input-len 32768 and
--max-kv-cache-capacity 40960. The input limit covers the prompt, including
a client’s system prompt and tools. The KV-cache capacity must also hold the
completion. Change either flag when a workload needs a different length.
AIPerf stays at 8,192 and 10,240: keep each trace inside that limit instead of
building a longer engine.
Add --enable-auto-tool-choice for Claude Code, OpenClaw, and Codex. The
server checks a bearer token only when --api-key is set.
Tool |
Endpoint |
|---|---|
|
|
|
|
|
|
|
|
|
|
|