Inkling-Small vs Qwen3.5-397B-A17B

Inkling-Small

A mixture-of-experts generalist with images and audio in, several tool calls a turn and per-request thinking effort. It needs about 205 GB for Thinking Machines' NVFP4 build on recent NVIDIA or AMD datacentre GPUs, and Tinker or DeepInfra host it. The main risks are a use policy that binds on download and can change, and no outside safety results.

Qwen3.5-397B-A17B

Alibaba's biggest open Qwen3.5, a cousin for hard coding, agent and vision work. Even at 4 bits it needs about 270 GB, so most agents will call it through Alibaba's Model Studio or one of the many providers that host it. The main risk is silence, with no safety results, no agent guidance and template reports without a reply.

Inkling-Small

1. Start the NVFP4 build with SGLang's recipe, sglang serve --trust-remote-code --model-path thinkingmachines/Inkling-Small-NVFP4 --tp 8 --enable-multimodal --reasoning-parser inkling --tool-call-parser inkling, on Blackwell for W4A4
2. Use the inkling tool and reasoning parsers in vLLM or SGLang, or Thinking Machines' tml-renderers package to render prompts and parse calls yourself
3. 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 and SGLang's cookbook list xhigh
4. Effort none conditions the model toward no reasoning but doesn't forbid it, and higher effort needs a larger max_tokens
5. Tool call arguments must reach the template as an object, not a JSON string, or it raises an error
6. Images should be 40 to 4096 px a side and audio 16 kHz WAV under two minutes, per the card
7. Use guided decoding or validate JSON yourself, since structured output isn't documented
8. Read the Model Acceptable Use Policy before deploying, it binds on download and bars some uses that are legal

Qwen3.5-397B-A17B

1. The card's server command is vllm serve Qwen/Qwen3.5-397B-A17B --port 8000 --tensor-parallel-size 8 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder, and the FP8 build halves the memory
2. It doesn't fit our 128 GB reference machine at any precision, so call a hosted endpoint instead (Model Studio's Qwen3.5-Plus, or a provider serving Qwen/Qwen3.5-397B-A17B)
3. On Model Studio pass "enable_thinking": false directly, while vLLM and SGLang take "chat_template_kwargs": {"enable_thinking": false}
4. 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
5. For multi-token prediction in vLLM add --speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":2}'
6. Keep the context at 128K or more if you rely on thinking, and turn on YaRN only past 262,144 tokens
7. Validate any JSON it returns, since structured output isn't documented