Nemotron 3.5 Lightning 30B A3B vs Qwen3.5-4B

Nemotron 3.5 Lightning 30B A3B

A small agent cousin, 3B active parameters, with trained parallel tool calls, a thinking switch and up to 1M tokens of context, under a permissive licence. It runs in about 23 GB from the GGUF or Ollama builds, and NVIDIA's NVFP4 build needs NVIDIA GPUs. The main risks are no published safety results and one answered question across two repositories.

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 3.5 Lightning 30B A3B

1. Quickest start is ollama run nemotron-3.5-lightning, about 19 to 25 GB of files depending on the platform
2. On the reference machine use llama-server -hf ggml-org/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF:Q4_0 --jinja, since the card's Q4_K_M tag doesn't exist there and NVFP4 needs NVIDIA GPUs
3. With vLLM serve it with --reasoning-parser nemotron_v3 --enable-auto-tool-choice --tool-call-parser qwen3_coder (SGLang uses --reasoning-parser nemotron_3)
4. Use temperature=1.0 and top_p=0.95, as the card says
5. Thinking is on by default. Pass chat_template_kwargs with enable_thinking false for short tasks
6. For coding agents pass force_nonempty_content: true in chat_template_kwargs, which vLLM's nemotron_v3 parser reads so a cut-off reasoning trace still returns content
7. Go past 262,144 tokens only with the card's flags (VLLM_ALLOW_LONG_MAX_MODEL_LEN=1, --max-model-len 1048576) and enough memory

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