Nemotron 3 Nano 30B A3B vs Spark-X2.6-4B

Nemotron 3 Nano 30B A3B

A fast hybrid Mamba-Transformer cousin with 3.5B active parameters for reasoning and tool calling over very long inputs, in every major local runtime. A 4-bit build needs about 22 to 23 GB. The main risks are that NVIDIA's documented vLLM path runs repository code, no safety results are published and NVIDIA hasn't answered reports since March.

Spark-X2.6-4B

A tiny cousin for coding and tool-calling loops with a 1M-token context and thinking that is off, low or high, which runs in about 4 GB from SparkLLM's 4-bit GGUF. Use llama.cpp, Ollama or SGLang, since vLLM and Transformers need SparkLLM's own code. The risks are no safety work or agent guidance and a thin, two-day-old card.

Nemotron 3 Nano 30B A3B

1. Start it with ollama run nemotron-3-nano:30b (18 GB MLX on a Mac, 24 GB q4_K_M elsewhere)
2. On the reference machine use BF16 with vLLM, after wget https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/resolve/main/nano_v3_reasoning_parser.py, then vllm serve nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 --max-model-len 262144 --trust-remote-code --enable-auto-tool-choice --tool-call-parser qwen3_coder --reasoning-parser-plugin nano_v3_reasoning_parser.py --reasoning-parser nano_v3
3. NVFP4 needs an NVIDIA Blackwell GPU, and FP8 isn't documented on gfx1151
4. Without the qwen3_coder tool parser and the nano_v3 reasoning parser, tool calls and reasoning come back as plain text
5. Card sampling is temperature 0.6 and top_p 0.95 for tool calling, temperature 1.0 and top_p 1.0 for reasoning
6. Thinking is on by default, so pass enable_thinking as false in chat_template_kwargs for short tasks
7. Set VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 and raise --max-model-len only when you need more than 262,144 tokens
8. In SGLang the main branch names the reasoning parser nemotron_3, not the card's nano_v3
9. Filter untrusted tool results before they reach the model, as NVIDIA warns prompt injection can lead to code execution
10. Validate JSON output yourself or use the runtime's guided decoding, since none is documented for the model

Spark-X2.6-4B

1. Quickest start is llama-server -hf XHToken/Spark-X2.6-4B-GGUF:Q4_K_M --jinja, which needs llama.cpp b10828 or later, or ollama run SparkLLM/Spark-X2.6-4B (8.2 GB, Ollama 0.34.1 or later)
2. For parsed tool calls serve it with SGLang 0.5.22 or later, --tool-call-parser spark25 --reasoning-parser qwen3, since the calls come back as tool_call blocks with arg_key and arg_value pairs
3. Thinking is on by default. Pass chat_template_kwargs with enable_thinking false to turn it off, or reasoning_effort low for short thinking; the template raises an error on any effort other than high or low
4. SparkLLM's recommended sampling is temperature 1.0, top_p 0.95 and top_k -1
5. The template always opens with its own default system line, with the effort instruction, and puts your system message after the tool list
6. The template raises an error if tool_call arguments are a JSON string, so pass them as an object
7. Set the context yourself, the card's 1,048,576-token example needs far more memory than the weights