Inkling vs Qwen3-4B-Instruct-2507

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.

Qwen3-4B-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.

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

Qwen3-4B-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