We’re running evaluations on nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B through the OpenAI-compatible endpoint (https://api.deepinfra.com/v1/openai/chat/completions). Completions stop at exactly 16,384 tokens even when max_tokens is set higher.
Reproduction (2026-10-27, 16:19:09 UTC):
- Model: nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B
- max_tokens: 40000, temperature: 0, reasoning_effort: “none”
- Prompt: “Write the integers from 1 to 20000, one per line, with no other text.”
- Result: HTTP 200, finish_reason: “length”, completion_tokens: 16384
- x-request-id: RRNmS1T38TQIsi1ufxeI0muX
The same thing happened on an earlier call at about 16:10 UTC, also with max_tokens 40000.
For comparison, we sent the identical request to nvidia/NVIDIA-Nemotron-3.5-Super-VL at the same time. It returned the full 40,000 tokens. The /models metadata lists max_tokens: 262144 for the Ultra model.