inkling 2 mini is the open-weights model a lot of developers are looking at this weekend. Thinking Machines Lab released it on 20 November 2026 under the Apache 2.0 license, and it is aimed at agentic and tool-use systems, coding assistants, chatbots and RAG apps (Hugging Face model card).
If you build automations for a small business, the useful question is simple: can you run it, and is it worth trying? Here is what the model card says, in plain terms.
The key facts
All of these come from the official model card:
- Size: 26.4B total parameters, 3.4B active per token. Because its middle layers loop three times, the card puts it at about 5.4B effective active parameters per token.
- Inputs and outputs: text and images in, text out.
- Context window: up to 1M tokens.
- Download: a single checkpoint of about 20GB, in mixed NVFP4 and MXFP8 precision.
- License: Apache 2.0, so commercial use and fine-tuning are allowed under its terms.
What hardware you need
The card lists three supported setups for local inference:
- One NVIDIA DGX Spark.
- Apple silicon with 32GB or more of memory (MacBook, Mac mini or Mac Studio).
- Other GPUs, with 32GB+ of VRAM recommended.
If your machine has less than that, you have two routes. Thinking Machines offers API access and an inference-only playground through its Tinker platform. And the community is already making smaller quantized versions: a Hacker News post titled "Inkling-2-Mini in 5.4 GB Memory" is getting attention (HN). Commenters there point out the official checkpoint is already trained for low precision, so re-quantizing it further may cost quality. Treat smaller community builds as experiments, not production.
Local setup options
The card lists support in Hugging Face, vLLM, SGLang, TokenSpeed, NVIDIA Dynamo, Unsloth (including GGUF), llama.cpp for GGUF builds, and Ollama and Exo Labs on Apple silicon. For most small teams, the easiest path is:
- On a Mac with 32GB+: try Ollama first.
- On a GPU server: vLLM gives you an OpenAI-style endpoint you can point n8n, Make or your own scripts at.
A quick fit check before you switch
The card's own benchmark table shows strong agent and tool-use scores against similar-size open models. But vendor benchmarks are not your workload. Before moving an automation over:
- Pick 20 real tasks from your current flow (emails to sort, tickets to tag, JSON to extract).
- Run them on your current model and on inkling 2 mini with the same prompts.
- Check tool calls and JSON output, not just the wording. Broken structure breaks automations.
- Time it on your hardware. A local model only saves money if it is fast enough for the job.
- Keep a fallback. Route anything that fails validation to your existing API model.
Why it matters
An Apache-licensed model that can run on a 32GB Mac gives small teams a real option for private, low-cost automations, such as handling customer data you would rather not send to a third-party API. It is worth an afternoon of testing.
This post is research-based, using the sources linked above. For more plain-English AI news, visit ainews.useaifix.com, and for help setting up AI automations for your business, see useaifix.com.
Top comments (0)