Kimi K3 vs Qwen3-4B-Instruct-2507

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.

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.

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

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