Inkling vs Qwen3.5-9B

Inkling

Thinking Machines' large multimodal generalist, stronger than Inkling-Small on factual recall by its own numbers. It needs about 710 GB of datacentre GPU memory even in NVFP4, so most agents will use a host such as Together AI or Tinker. The main risks are a use policy that binds on download and can change, and no outside safety results.

Qwen3.5-9B

A small dense vision-language cousin with trained tool calling, switchable thinking and a 262,144-token context, for local agent subtasks. Ollama's 4-bit build needs about 8 GB and the BF16 files about 24 GB, in every major runtime. The main risk is neglect, with no file changes since March, no replies from Qwen to reports and no safety results.

Inkling

1. Unless you have a multi-GPU Blackwell or H200 node, call it through Together AI, Fireworks, DeepInfra or Tinker rather than starting it
2. On 8 B200 GPUs, SGLang's recipe is sglang serve --trust-remote-code --model-path thinkingmachines/Inkling-NVFP4 --tp 8 --quantization modelopt_fp4 --enable-multimodal --reasoning-parser inkling --tool-call-parser inkling
3. Use the inkling tool parser, not Hermes, and the inkling reasoning parser
4. Pass reasoning_effort as none, minimal, low, medium, high (the default) or max, or a number from 0 to 0.99. The repository's template has no xhigh preset and raises an error for it, though Thinking Machines' docs list xhigh
5. Tool call arguments must reach the template as an object, not a JSON string, or it raises an error
6. Audio should be 16 kHz WAV, ideally under 20 minutes, and images 40 to 4096 px a side, per the card
7. Read the Model Acceptable Use Policy before deploying, it binds on download and bars some uses that are legal

Qwen3.5-9B

1. Smallest start is ollama run qwen3.5:9b (6.6 GB, q4_k_m)
2. On vLLM run vllm serve Qwen/Qwen3.5-9B --port 8000 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder, or tool calls come back as text
3. Tool calls use the XML style <tool_call><function=name><parameter=x>, not Hermes JSON, so use the qwen3_coder parser and not hermes
4. Thinking is on by default; pass chat_template_kwargs with enable_thinking false for short tasks, since /nothink isn't supported
5. There's no generation_config.json, so pass sampling yourself, temperature 1.0, top_p 0.95, top_k 20, presence_penalty 1.5 when thinking, or 0.7, 0.8, 20, 1.5 without
6. Keep at least 128K of context when thinking, and turn on YaRN only beyond 262,144 tokens
7. Validate JSON output yourself or use the runtime's guided decoding