Ling-3.0-tiny vs Qwen3.5-122B-A10B

Ling-3.0-tiny

A small, quick hybrid reasoning cousin for local agent work, tool calls and coding, under MIT. It runs in about 6 GB from inclusionAI's 4-bit GGUF in released Transformers, vLLM, SGLang and llama.cpp, with no host. The main risks are a card that doesn't match the files in places, the vLLM route running repository code, and no model safety results.

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.

Ling-3.0-tiny

1. Start it with llama-server -hf inclusionAI/Ling-3.0-tiny-GGUF:Q4_K_M --jinja -c 32768, which needs llama.cpp b11772 or a recent release with bailingmoe3
2. With vLLM 0.32.0, vllm serve inclusionAI/Ling-3.0-tiny --trust-remote-code --enable-prefix-caching --mamba-cache-mode align --enable-auto-tool-choice --tool-call-parser ling3 --reasoning-parser ling3, which runs the repository's config code
3. With SGLang 0.5.22, its cookbook uses --reasoning-parser deepseek-r1 --tool-call-parser glm45 for this model, and drop the card's NEXTN flags since tiny has no MTP layer
4. Thinking is on by default, pass chat_template_kwargs with enable_thinking false for short tasks
5. Sampling per the card is temperature 1.0, top_p 0.95 and top_k 20
6. The template keeps earlier reasoning in the prompt, which grows long agent sessions
7. Validate JSON output yourself or use the runtime's guided decoding, since structured output isn't documented
8. For a guardrail, inclusionAI's Ling-3.0-tiny-singprobe scores intent, unsafe output and hallucination per token, but only through SGLang and vLLM integration branches

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