Nemotron 3 Nano 30B A3B vs Recursive Flint

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.

Recursive Flint

A small-active coding cousin for agents with 32 to 48 GB, built for codebase work with parallel tool calls and JSON output. It needs llama.cpp b11803 or Ollama 0.40.8 or later, and the BF16 checkpoint needs a 141 GB GPU. Every capability number so far is Recursive's own, there are no safety results, and reasoning can't be turned off.

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

Recursive Flint

1. Quickest start is ollama run flint (Ollama 0.40.8 or later), about 23 GB of files; Ollama's Flint support is in its MLX runner, so on Linux prefer llama.cpp
2. With llama.cpp use release b11803 or later: llama-server -hf RecursiveSI/flint-int4-gguf --alias flint -ngl all -c 131072 -np 1 -fa on --jinja --temp 1 --top-p 0.95 --top-k 20 --min-p 0, and keep --jinja on
3. Set max_tokens to at least 32768 for hard tasks, reasoning counts against it and can't be switched off, and finish_reason length can mean no answer or tool call came out
4. Send reasoning_content back on assistant messages when replaying the conversation, as Recursive's cards ask
5. Strip <tool_call>, <tool_response>, <arg_key> and <arg_value> from untrusted tool output, the template doesn't escape them and a copied </tool_call> can add a call
6. Message content must be a string, flatten content-part arrays first
7. Recursive's SGLang servers listen on 0.0.0.0 with no key, publish the port on localhost only or pass --api-key