Request Shapes and Options

🧬 Request shapes and options

The proxy understands two payload styles:

• Ollama-style: model parameters belong in options. Use this for /api/* endpoints.
• OpenAI-style passthrough: send OpenAI-compatible JSON to /v1/*, forwarded untouched.

For Ollama endpoints, put fields such as temperature, num_predict, max_tokens, logit_bias, and structured format values inside the options object (or top-level format). Set "format": "json" for quick JSON, or pass a JSON Schema object (also accepted inside options.format) to use LM Studio's structured output enforcement.

Options that go inside options

Ollama option	LM Studio parameter	Notes
temperature, top_p	Same name	Direct passthrough
top_k	top_k	Direct passthrough
min_p	min_p	Direct passthrough
presence_penalty	presence_penalty	Direct passthrough
frequency_penalty	frequency_penalty	Direct passthrough
repeat_penalty	repeat_penalty / frequency_penalty	Mapped depending on what is already set
max_tokens / num_predict	max_tokens	Picks whichever you set; max_tokens takes priority
num_ctx	context_length	Reloads the model at the requested context length before inference (LM Studio treats this as a load-time setting). No-op when absent/zero or already loaded at that size. Clamped to the model's max. Two concurrent requests with different num_ctx to the same model can race. When absent, falls back to --default-context-length / OLLAMA_CONTEXT_LENGTH if set. Also honored on /api/embed
logit_bias	logit_bias	Accepts JSON object or map notation
system (in options)	system	Injected as LM Studio system prompt
stop, seed	Same name	Direct passthrough
truncate	truncate	Direct passthrough; defaults to true on /api/embed when omitted (matches Ollama) so overlong inputs truncate instead of erroring
dimensions	dimensions	Direct passthrough (embeddings)

Accepted but ignored

Ollama-only sampling/runtime options with no LM Studio equivalent (e.g. mirostat*, tfs_z, typical_p, repeat_last_n, num_keep, num_gpu, num_thread, numa) are accepted without error and dropped, surfaced once per request in a warn-log so they are not silently swallowed.

draft_num_predict (max speculative draft tokens per step) falls here too: LM Studio configures speculative decoding at model-load time via a draft model, with no per-request knob in its API, so the option is dropped — a no-op, matching Ollama itself when no draft model is loaded.

Fields that go at the top level

Ollama field	LM Studio parameter	Notes
think / reasoning_effort	reasoning	true→"on", false→"off", "none"→"off"; levels low|medium|high|on|off pass through; reasoning_effort is an alias used only when think is absent. When think is omitted and the model is reasoning-capable (LM Studio reports a reasoning capability), defaults to "on" to match Ollama; explicit think:false always wins
logprobs, top_logprobs	Same name	Direct passthrough
suffix	suffix	Forwarded on non-vision generate requests only
raw	none	Disables system-prompt injection in generate requests
keep_alive	ttl	Seconds (int) or duration string ("5m"); 0 unloads the model immediately
tool_choice	tool_choice	Forwarded on /api/chat (OpenAI-compat path) when tools is also present. Not forwarded without tools, and not on the --use-native-chat path (the native schema has no slot; the response carries a warning field naming the dropped fields)
integrations	integrations	Native path only (--use-native-chat, or --native-chat-streaming on a streaming request). Array of MCP tool specs forwarded verbatim. See MCP Integrations.