Inkling-Small vs Qwen3-Coder-Next

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-Coder-Next

A fast coding cousin for agent harnesses, with 3B active parameters and tool calling trained across many scaffold formats for Claude Code, Cline or Qwen Code loops. Alibaba's 4-bit GGUF needs about 58 GB, fitting one 64 GB machine, and two providers host it. The main risk is neglect, with no change since February, no replies and no safety results.

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-Coder-Next

1. Quickest start is ollama run qwen3-coder-next (Q4_K_M, 52 GB of files)
2. On our reference machine Alibaba's Q8_0 GGUF (84.8 GB, about 102 GB in use) also fits, run with llama-server -m Qwen3-Coder-Next-Q8_0/Qwen3-Coder-Next-Q8_0-00001-of-00004.gguf --jinja -ngl 99 -fa on -c 131072
3. Serve with vllm serve Qwen/Qwen3-Coder-Next --enable-auto-tool-choice --tool-call-parser qwen3_coder (vLLM 0.15.0 or later), using the built-in parser rather than the repository's qwen3coder_tool_parser_vllm.py, which fails to import on newer vLLM per discussion 36
4. The card's commands use tensor parallel 2 though the text says 4 GPUs, and it suggests cutting the context to 32,768 if the server fails to start
5. There's no thinking mode, so don't pass enable_thinking or wait for reasoning_content
6. Sampling per the card is temperature 1.0, top_p 0.95 and top_k 40
7. The GGUF card's llama.cpp example path omits the quant from the file name; the real files are named like Qwen3-Coder-Next-Q4_K_M-00001-of-00004.gguf
8. Validate any JSON it returns, since structured output isn't documented