Kimi K3 vs Nemotron 3 Super 120B A12B

Kimi K3

Kimi K3 is a 2.8T multimodal agent model with a 1M-token context, parallel tool calls and per-request effort, for long coding and research sessions. Its 1,561 GB of files need a multi-node GPU cluster, so most agents will call a hosted copy. The main risks are Moonshot's Python running at load and prompt time, and no answers in its repository.

Nemotron 3 Super 120B A12B

A long-context agent cousin with trained parallel tool calls, a reasoning switch and a full account of its training data, under a free commercial licence. It needs about 100 GB of GPU memory, a 128 GB unified machine or Blackwell hardware for NVIDIA's NVFP4. The main risks are serving commands that run repository code and no published safety results.

Kimi K3

1. Start it with SGLang on a node of 8 B300s, sglang serve --trust-remote-code --model-path moonshotai/Kimi-K3 --tp-size 8 --context-length 65536 --mem-fraction-static 0.82 --reasoning-parser kimi_k3 --tool-call-parser kimi_k3 --port 30000
2. Use the kimi_k3 tool-call and reasoning parsers in vLLM or SGLang, and validate tool-call arguments against the schema, since vLLM's recipe says K3 sometimes emits a format its parser doesn't expect
3. Pass the whole assistant message back every turn, reasoning_content and tool_calls included, or output becomes unstable
4. Thinking is always on; set reasoning_effort to low or high when max is too slow
5. Sampling per the card is temperature 1.0, with top-p 0.95 for single-step tasks and 1.0 for agentic ones
6. Don't hand an ongoing session from another model to K3, and state boundaries in the system prompt, since Moonshot warns it can act on its own when intent is unclear
7. Without the hardware, call it through Moonshot's API or an inference provider listed on its Hugging Face page

Nemotron 3 Super 120B A12B

1. On the reference machine start it with ollama run nemotron-3-super, the 87 GB q4_K_M build, the only one that fits (about 104 GB with the margin)
2. Raise the Strix Halo GPU memory limit first with amd-ttm --set 100, since the default of about 62.8 GiB is far below what it needs
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 for every task, as the card says
5. Thinking is on by default. Pass enable_thinking: false or low_effort: true in chat_template_kwargs for short tasks
6. For coding agents pass force_nonempty_content: true in chat_template_kwargs, so a cut-off reasoning trace still returns content
7. Raise context past 262,144 only with the card's flags (VLLM_ALLOW_LONG_MAX_MODEL_LEN=1, --max-model-len 1048576)