All Nebius Nemotron models incompatible with NemoClaw Option 3 (reasoning-only, no non-reasoning mode)

[troopdegen]

Severity: High — blocks the natural NVIDIA-on-Nebius integration path for NemoClaw users

Discovered: 2026-05-06 during tutorial development (Nebius Token Factory + NemoClaw integration)

Description

All three NVIDIA Nemotron models available on Nebius Token Factory are reasoning models:

• Llama-3_1-Nemotron-Ultra-253B-v1
• nemotron-3-super-120b-a12b
• NVIDIA-Nemotron-3-Nano-30B-A3B

They return responses in reasoning_content with an empty content field. NemoClaw's Option 3 (OpenAI-compatible endpoint) hardcodes NEMOCLAW_REASONING=false and reads only choices[0].message.content, causing:

1. Smoke check failure during onboarding — the 32-token budget is exhausted by the reasoning trace; content is null/empty
2. 400 errors on tool calls — reasoning models do not support tool calls via the OpenAI-compatible wrapper

Nebius is listed as a supported inference provider in the NemoClaw docs, and NVIDIA Nemotron models are a natural first choice for developers using NemoClaw on Nebius. The combination is currently broken with no documentation warning.

Workaround

Use a non-reasoning model — deepseek-ai/DeepSeek-V3.2, meta-llama/Llama-3.3-70B-Instruct, or NousResearch/Hermes-4-70B.

Expected fix

Either expose a non-reasoning inference mode for Nemotron models via the Nebius OpenAI-compatible endpoint, or document the incompatibility clearly in the NemoClaw integration guide on the Nebius side.

Note: a companion issue has been filed with NVIDIA/NemoClaw requesting runtime NEMOCLAW_REASONING configuration for Option 3 providers.