Nemotron 3 Ultra 550B A55B vs Qwen3-4B-Instruct-2507

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.

Qwen3-4B-Instruct-2507

A small, fast cousin for instruction following, short tool-calling jobs and long documents on modest hardware, with no thinking step. It runs in about 3 GB at 4 bits on every common runtime, and two providers host it. The main risk is neglect, with no file changes in over a year, no answers from Alibaba and no safety results.

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

Qwen3-4B-Instruct-2507

1. Quickest start is ollama run qwen3:4b-instruct, 2.5 GB. Plain qwen3:4b is the Thinking-2507 edition, not this one
2. On our reference machine use BF16 with vllm serve Qwen/Qwen3-4B-Instruct-2507 --max-model-len 262144 --enable-auto-tool-choice --tool-call-parser hermes
3. Use the hermes parser in vLLM or qwen25 in SGLang, the template emits JSON inside tool_call tags
4. Don't pass enable_thinking, this edition never thinks and the card says the flag isn't needed
5. Sampling per the card is temperature 0.7, top_p 0.8, top_k 20 and min_p 0, with presence_penalty up to 2 against repetition
6. Drop the context to 32,768 tokens if memory runs out, as the card suggests
7. Validate JSON yourself or use the runtime's guided decoding, since structured output isn't documented