Kimi K3 vs Qwen3.8-27B

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.

Qwen3.8-27B

A dense cousin for coding and agent work with image and video input that runs almost anywhere. Ollama's 4-bit build needs about 22 GB, and five providers host it for agents without the hardware. The main risk is that nobody at Alibaba answers, there are no safety results or agent guidance, and the tool format is only in the template.

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

Qwen3.8-27B

1. Quickest start is ollama run qwen3.8, about 18 GB of files
2. On our reference machine use BF16 with vllm serve Qwen/Qwen3.8-27B --max-model-len 131072 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder, since FP8 on gfx1151 isn't documented
3. Use the qwen3_coder tool parser, SGLang's cookbook says the Hermes parser leaves its calls unparsed
4. Thinking is on by default, pass chat_template_kwargs with enable_thinking false for short tasks, and read the reasoning from reasoning_content
5. reasoning_effort takes xhigh (default), medium or low, and any other value makes the template raise an error
6. Sampling per the card is temperature 1.0 and top_p 0.95 when thinking, temperature 0.7, top_p 0.8 and presence_penalty 1.5 without
7. Preserved thinking keeps earlier reasoning in the context unless preserve_thinking is false, which grows long agent sessions
8. Turn on YaRN only beyond 262,144 tokens, the card warns static YaRN can cost quality on short texts
9. Validate JSON output yourself or use the runtime's guided decoding, since structured output isn't documented