Nemotron 3 Ultra 550B A55B vs Qwen3.5-4B

Nemotron 3 Ultra 550B A55B

NVIDIA's frontier-scale agent cousin with trained parallel tool calls, up to a 1M-token context and training against prompt injection through tool output. It needs at least four Blackwell GPUs for the 352 GB NVFP4 build, so most agents will use NVIDIA's API or a provider. The main risks are serving commands that run repository code and no published safety results.

Qwen3.5-4B

A small vision-language cousin for tool calls, light agent steps and image or video reading on almost any machine, with switchable thinking. Ollama's 3.3 GB build or the 9.3 GB BF16 files run in every major runtime under plain Apache 2.0. The main risks are no safety results, no structured-output guarantee and no answers from Alibaba since March.

Nemotron 3 Ultra 550B A55B

1. It doesn't fit the 128 GB reference machine at any precision. Call NVIDIA's API (model nvidia/nemotron-3-ultra-550b-a55b) or a provider instead
2. With vLLM serve the NVFP4 build with --trust-remote-code --reasoning-parser nemotron_v3 --enable-auto-tool-choice --tool-call-parser qwen3_coder on 4 × B200 or more
3. With SGLang use --reasoning-parser nemotron_3, not the BF16 card's nemotron_v3, which SGLang doesn't register
4. On SGLang with tools, pass chat_template_kwargs with enable_thinking true and force_nonempty_content true, as the card says, or reasoning and calls won't both parse
5. Thinking is on by default. Pass enable_thinking: false for short tasks or medium_effort: true to cut reasoning tokens
6. Use temperature=1.0 and top_p=0.95, as the generation config sets

Qwen3.5-4B

1. Smallest start is ollama run qwen3.5:4b (Q4_K_M, about 4 GB with context)
2. Full precision vllm serve Qwen/Qwen3.5-4B --port 8000 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder
3. Turn thinking off with "chat_template_kwargs": {"enable_thinking": false} in the request body, since it's on by default and /think and /nothink aren't supported
4. There's no generation_config.json, so pass the card's sampling yourself, for example temperature=0.7, top_p=0.8, top_k=20, presence_penalty=1.5 without thinking
5. Keep the context at 128K or more if you rely on thinking, and turn on YaRN only past 262,144 tokens
6. Validate any JSON it returns, since structured output isn't documented