Kimi K3 vs Qwen3.5-122B-A10B

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.

Qwen3.5-122B-A10B

A strong mid-size cousin for coding, agent steps and reading images or video, with thinking you can switch off per request. Alibaba's 4-bit build needs about 95 GB, which fits one 128 GB unified-memory machine, and several providers host it. The main risk is silence, with no safety results, agent guidance or answers from Alibaba.

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

Qwen3.5-122B-A10B

1. Quickest start on a 128 GB machine is ollama run qwen3.5:122b (Q4_K_M, 81 GB of files, about 97 GB with context)
2. Alibaba's 4-bit build serves with vllm serve Qwen/Qwen3.5-122B-A10B-GPTQ-Int4 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder --quantization moe_wna16, though that kernel isn't documented for AMD gfx1151
3. Use the qwen3_coder tool parser and the qwen3 reasoning parser, as the card's commands do
4. Turn thinking off with "chat_template_kwargs": {"enable_thinking": false} in the request body, since it's on by default and /think and /nothink aren't supported
5. Sampling per the card is temperature 0.6, top_p 0.95 and top_k 20 when thinking, and temperature 0.7, top_p 0.8, top_k 20 and presence_penalty 1.5 without
6. Keep the context at 128K or more if you rely on thinking, and turn on YaRN only past 262,144 tokens
7. Add --language-model-only in vLLM to skip the vision encoder and free memory for the KV cache when you only need text
8. Validate any JSON it returns, since structured output isn't documented