Limits and context

Context per model

The room’s context is the most tokens one request can hold: prompt plus answer.

Model	Default	Most with ?ctx=
Qwen3 1.7B	16384 (8192 when the room is short of memory for 16K)	16384
Qwen3.8 27B	16384	32768
Qwen3.6 35B MoE	32768	65536

The host sets it by opening the room page with ?ctx= before starting the model, for example pooled.run/room?ctx=65536. The value is rounded to a multiple of 256, at least 2048.

The bridge reports it in the banner, as ctx in GET /health and as max_model_len in GET /v1/models. Give it to coding agents so they compact in time: see context settings.

Prompt and answer length

• Prompt. It must leave at least 32 tokens of the context free. A longer one is a 400 with code context_length_exceeded. The bridge refuses early, without asking the room, when the text is over 8 characters per token of context.
• Answer. max_tokens (or max_completion_tokens, max_output_tokens) defaults to 16384 on Chat Completions and Responses. Messages requires it. Above 65536 it is capped, not refused. The answer always stops when the context is full.

When the context fills mid-answer: Chat Completions returns finish_reason: "length", Responses status: "incomplete", Messages stop_reason: "model_context_window_exceeded".

Request size

Limit	Value	Over it
HTTP body	4 MiB	413
The request as sent to the room	3.5 MB	413
Text in the whole request	1,500,000 characters	400
Messages	1000	400
stop / stop_sequences	4 strings, 64 characters each	400
Tools	1024	400
Connections at once	64

Tool and schema limits are on Tool calling and Structured output.

Queues

The room runs one generation at a time for its chat, Code mode and API clients.

1. In the bridge. One request goes to the room at a time, and up to 8 more wait (--max-queue).
2. In the room. The host queues at most 10 questions, 2 per member.

When either is full, the request gets 429 (OpenAI) or 529 (Anthropic) with Retry-After: 5.

While a request waits or the room reads a long prompt, streams get a keep-alive after 10 s of silence: an SSE comment on Chat Completions, a ping on Messages, response.in_progress on Responses. A non-streamed request sends nothing until it is done, so use streaming if your client has a short read timeout.

Timeouts

• If the room sends nothing for a request for 5 minutes, the bridge returns 504. While the host holds it in its own queue, the limit is 30 minutes.
• There is no limit on total time. A long prompt on a room of phones can take minutes before the first token.
• Closing the connection stops the request in the room.

Every status code is on HTTP endpoints.