Thinking Machines releases open-weight Inkling 2 Mini

Online Tech Tips Staff · Edited by Aseem Kishore
6 min read

Online Tech Tips is reader-supported. We may earn a commission when you buy through links on our site. Learn more.

On November 20, 2026, Thinking Machines announced Inkling 2 Mini, an open-weight model whose full weights are downloadable from Hugging Face under the Apache 2.0 license. The company pitches it as small enough to run on a laptop, but you’ll need software and hardware that support it.

That’s appealing if you’re tired of sending every question to an online chatbot. The evidence suggests a promising local option, but I’d keep your existing service for now.

What Happened: Inkling 2 Mini Becomes Downloadable

Inkling 2 Mini accepts text and image inputs and replies in text. Its stated context window holds up to one million tokens. Tokens are the small pieces of information an AI processes. Think of context as a working desk with room for your material. Your computer still needs enough memory.

Image input means you can give the model pictures to process. That doesn’t mean it can generate images or answer every question about a picture correctly.

Thinking Machines’ announcement says open-source inference is available on SGLang, vLLM, Ollama, Unsloth, NVIDIA Dynamo, TokenSpeed and Llama App, with quantized GGUF versions via Unsloth and llama.cpp, and the Hugging Face article adds MLX support for Apple silicon. Those are runtimes, not finished chat apps, so you’ll still need a front end you like.

Capture the official release article title, November 20, 2026 publication date, and introductory text describing Inkling 2 Mini's capabilities.

Why the Parameter Counts Differ

Parameters are the learned number settings inside a model. The official model card lists three figures:

  • 26.4 billion total parameters: Everything you download and store.
  • 3.4 billion active parameters: The weights used for each token in a single pass.
  • 5.4 billion effective active parameters: The per-token figure once the looped layers run three times, which is the number Artificial Analysis uses in its comparisons.

Hugging Face explains that the 12 middle layers of its 24 run three times for every token. This reuse increases the active parameter count without adding new layers to the model. It’s like rereading selected pages without adding more pages to the book.

The loop doesn’t make the model lighter to run, since it still makes 48 layer passes per token, but Artificial Analysis found it needed less total compute than Qwen3.8 27B on its own test suite. You’ll still need to plan memory for the full model. Don’t use the active parameter count alone as a guide to its memory needs.

How It Stacks Up

Independent Artificial Analysis testing compares Inkling 2 Mini with other models on its Intelligence Index.

Every figure below comes from that test. Task costs use hosted prices. They don’t include local electricity or hardware costs.

ModelIntelligence IndexReported hosted cost per task
Inkling 2 Mini20$0.59
Qwen3.8 27B (xhigh)Comparable to Inkling$3.53
Nanbeige4.3-33BComparable to Inkling–

Artificial Analysis reports task times of 9.0 minutes for Inkling versus 38.2 minutes for Qwen. It also estimates computing work by multiplying active parameters by tokens. That estimate favors Inkling by about 2.8 times against Qwen.

Separately, Inkling uses roughly six times fewer active parameters than Nanbeige4.3-33B. These results apply to the evaluator’s tests. They don’t measure laptop speed.

Separately, Hugging Face highlights benchmark results favoring Mini against its chosen comparison models. That’s a vendor-reported comparison. Neither that claim nor the Artificial Analysis results prove it beats the latest Claude, OpenAI, or Google Gemini models.

Coding and agent tests also don’t directly measure how helpful a chatbot feels in everyday use. Inkling’s downloadable weights, rather than a broad value claim, are the clearest reason to consider it over a hosted competitor.

Hosted Prices and Local Costs

Artificial Analysis says its task-cost figures use Thinking Machines’ first-party pricing for hosted Inkling 2 Mini, which the company offers through its Tinker platform:

  • Input: $0.15 per million tokens.
  • Output: $0.50 per million tokens.
  • Cache hits: $0.015 per million tokens.

Those are Thinking Machines’ launch rates as reported by Artificial Analysis. Third-party hosts serving the open weights set their own prices, so check the rate with whichever provider you plan to use.

These are per-token usage rates; Artificial Analysis doesn’t describe a subscription plan, minimums or extra platform fees, so check Tinker’s terms before you budget. Local use has its own costs, including hardware, electricity, and any application charges. A cheaper token rate alone won’t tell you which option costs less overall.

What Hardware You’ll Need

Thinking Machines puts the model’s total memory footprint at about 20GB. The model card lists support for Apple silicon Macs with 32GB or more of memory and NVIDIA’s DGX Spark, and says other GPUs and platforms are possible with 32GB or more of VRAM recommended. For Macs, the announcement names Ollama on systems with at least 32GB of RAM, and the Hugging Face article covers MLX and Llama App.

Windows PCs aren’t named specifically, so a Windows setup depends on a compatible GPU and a runtime such as llama.cpp or Ollama that lists Inkling 2 Mini support. Neither the announcement nor the model card mentions phone support.

If you just want to try it in a browser, Thinking Machines offers chat on its Tinker Playground. The Hugging Face repository is where you’ll find the model files.

The Reaction

In its launch-day write-up, the Hugging Face team that added library support said its own informal vibe tests found the model at least comparable with Qwen 3.8 27B, while noting that more validation is needed. That’s Hugging Face’s early assessment rather than broad, independent user feedback, which is still forming a day after release.

Thinking Machines itself flags weak spots: it says Inkling 2 Mini isn’t the strongest small model for multi-turn chat or instructions with many constraints, scoring 50.0% on MultiChallenge and trailing Qwen3.8-27B there.

Our Take

Verdict: Inkling 2 Mini deserves attention if you want local use and can handle the software setup. I’d keep an existing hosted chatbot if convenience matters most.

Running the weights locally can keep prompts on your computer, but the app you wrap around them may still have its own network features. Check what usage data it sends, its cloud features, and any connected tools before trusting it with sensitive files.

Wrapping Up

Before downloading, check one app’s model support, hardware guidance, and network behavior. I’d hold off on buying a new laptop or external SSD for Inkling alone. First, choose a setup you can repeat and check that its speed meets your needs.