Inkling vs Inkling Thinking: Which Mode Should You Use?
Thinking Machines released Inkling in two forms on NanoGPT: Inkling for direct answers and Inkling Thinking for work that benefits from more reasoning.
They are not two different models competing against each other. They are two ways to use the same model.
Short version:
- Use Inkling for quick questions, drafting, summaries, extraction, and straightforward image or audio analysis.
- Use Inkling Thinking for difficult coding, planning, tool use, forecasting, complicated instructions, and problems where the first answer is unlikely to be the best one.
- Start Thinking at medium effort. Move higher only when the task is hard enough to justify more waiting and a longer answer.
What is Inkling?
Inkling is an open-weights model from Thinking Machines. The company trained it from scratch and published the weights under the permissive Apache 2.0 license.
It is a mixture-of-experts model with 975 billion parameters in total and 41 billion active at a time. In plain terms, it has a very large pool of learned capacity but uses only part of that pool for each piece of a response. That helps balance capability with the amount of computation required.
For most users, the key details are:
- Text, image, and audio input
- Text output
- About a 1-million-token context window
- Up to 32,768 output tokens
- Tool calling and structured outputs on NanoGPT
- A direct mode and a controllable thinking mode
Thinking Machines is unusually direct about the model's position. Its launch post says Inkling is not the strongest model available overall, whether open or closed. The pitch is breadth: a generalist that combines reasoning, coding, tool use, multimodal input, long context, and open weights in one model.
Inkling and Inkling Thinking at a glance
| Inkling | Inkling Thinking | |
|---|---|---|
| Response style | Direct answer | Uses an extra reasoning budget before answering |
| Best for | Fast, clear, routine work | Difficult or multi-step work |
| Reasoning effort | No user-selectable effort setting | Minimal, low, medium, high, or xhigh |
| NanoGPT default | Direct answer | Medium effort |
| Inputs | Text, images, audio | Text, images, audio |
| Context | About 1 million tokens | About 1 million tokens |
The difference is how much work the model does before settling on its answer. Thinking is a dial, not a separate model. More effort can improve a hard answer, but it can also mean a longer wait and more generated tokens. For easy tasks, extra reasoning may add words without adding value.
Independent xhigh testing shows strong results with verbose output
Artificial Analysis tested the thinking-enabled Inkling configuration at xhigh reasoning effort. When we checked on July 17, 2026, it scored 41 on the Artificial Analysis Intelligence Index and ranked 10th out of 97 models on its comparison page that day.
That is a strong independent result for a newly released open-weights model. The same test also shows the tradeoffs:
- It generated visible output at about 72.7 tokens per second, above the 61.2-token median shown on the page. That is generation speed, not a guarantee that an xhigh answer will feel fast from start to finish.
- Across the full evaluation suite, it produced 130 million total output tokens, compared with a 92-million average for the models in its comparison class.
- Artificial Analysis described it as faster than average but somewhat verbose.
Those measurements are for xhigh, not ordinary direct Inkling or NanoGPT's medium Thinking default. They show what the model can do with a large reasoning budget, but not what every conversation will feel like.
The verbosity result matters. A model can have reasonable per-token pricing and still use more tokens than expected to complete a task. If Inkling Thinking starts over-explaining, lowering the effort or giving a clear length limit may help more than switching models.
What Thinking Machines reports
Thinking Machines published first-party results across reasoning, coding, tool use, instruction following, vision, and audio. The scores below show high-effort benchmark performance, not what direct Inkling or NanoGPT's medium Thinking default will do on every task:
| Benchmark | Inkling at effort 0.99, near maximum | What it measures |
|---|---|---|
| SWE-Bench Verified | 77.6% | Solving real software issues |
| Terminal-Bench 2.1 | 63.8% | Completing practical terminal tasks |
| IFBench | 79.8% | Following detailed instructions |
| MMMU Pro | 73.5% | Visual and academic reasoning |
| VoiceBench | 91.4% | Understanding spoken input |
These results come from Thinking Machines' own setup. Coding scores can also depend on the tools and surrounding setup, so they should not be treated as a guarantee for every prompt.
The broader table is more informative than any single score. Inkling is competitive across many categories, but does not lead all of them. It performs particularly well on instruction following and several agentic tasks, while stronger proprietary models and some other open models remain ahead on parts of advanced reasoning and coding.
Thinking Machines also reports an interesting efficiency result: on Terminal-Bench 2.1, Inkling matched the score of Nvidia's Nemotron 3 Ultra while using roughly one-third as many generated tokens. That is a company-reported comparison rather than an independent finding, but it is consistent with the idea behind effort control: use only as much thinking as the task needs.
When to use direct Inkling
Direct Inkling is the better default when the task is clear and speed matters.
Good examples include:
- Summarizing a document or conversation
- Rewriting text for tone or clarity
- Extracting names, dates, actions, or structured fields
- Classifying or tagging content
- Answering a focused question from supplied material
- Describing an image or identifying details in it
- Summarizing or discussing an audio recording
- Producing a quick first draft
- Simple coding questions with a clear answer
Direct mode is also useful when you expect to iterate. Three fast drafts can be more useful than one heavily reasoned draft, especially when the task depends on your taste rather than a single correct answer.
A direct prompt can stay simple:
Summarize the attached report in five bullets.
Separate confirmed findings from recommendations.
Keep the answer under 250 words.
Do not use Thinking by habit. If the direct model already gives a reliable answer, extra reasoning is only extra overhead.
When to use Inkling Thinking
Thinking mode makes more sense when the model needs to compare possibilities, recover from a false start, or plan several steps before answering.
It is worth using for:
- Debugging a problem with several plausible causes
- Planning a software change across multiple files
- Following a long set of constraints that may conflict
- Comparing contracts, policies, or research papers
- Forecasting with uncertain or incomplete evidence
- Building and checking a multi-step plan
- Tool-using workflows where an early mistake can affect later steps
- Visual or audio questions that require inference rather than description
- Reviewing a large amount of context before reaching a conclusion
For difficult work, ask for the result you need rather than simply asking the model to “think harder”:
Review these three implementation options.
For each one, identify the main benefit, failure risk, and maintenance cost.
Check the constraints against the attached specification.
Recommend one option, explain what evidence could change the recommendation,
and keep the final answer under 800 words.
That prompt gives the reasoning a job and still puts a limit on the final response.
Choosing a reasoning effort
Inkling Thinking supports five levels on NanoGPT:
- Minimal: A quick check before answering
- Low: Light comparison or straightforward planning
- Medium: The best starting point for most difficult tasks
- High: Hard coding, analysis, or multi-step work
- Xhigh: The largest reasoning budget for unusually difficult problems
The exact difference will vary by prompt. Higher effort does not guarantee a better answer, and xhigh should not be the automatic choice just because it produced the published benchmark scores.
A practical workflow is:
- Try direct Inkling for a well-defined task.
- Move to Thinking at medium if the answer misses interactions, constraints, or edge cases.
- Move to high or xhigh only when you can explain what the lower setting failed to do.
- Compare the result, not the length. A longer answer is not necessarily a better one.
If you would want a person to pause, check assumptions, and compare several paths before answering, use Thinking. If you would want them to answer immediately and let you react, use direct Inkling.
How useful is the 1M-token context window?
A context window of roughly one million tokens can hold a large amount of text, such as many documents, an extended conversation, or parts of a codebase. Images and audio are supported too, but their practical limits depend on the files and how they are processed.
Capacity is not the same as perfect recall. Relevant details can still be overlooked when they are buried in a very large input, especially when the request itself is vague.
For long-context work:
- Explain which documents or sections matter most.
- Ask the model to create an index or summary before drawing conclusions.
- Separate evidence gathering from the final recommendation.
- Request citations to filenames, headings, or quoted passages where possible.
- Keep a short working summary during a long task instead of relying only on old chat history.
The large window helps most when the material actually belongs together. It is not a reason to include unrelated files “just in case.”
What open weights changes
Inkling's weights are available to download under Apache 2.0, including for commercial use. That gives researchers and companies more freedom to download, inspect, customize, and deploy the weights than a closed API allows. It does not disclose the training data or make a model of this size cheap to host.
Open weights does not mean easy to run on a normal computer. A 975-billion-parameter model is still extremely large, even though only 41 billion parameters are active for each token. For most people, hosted access will remain more practical.
The weights are actually downloadable today, not promised for later.
The bottom line
Inkling is not trying to win every benchmark. It is a broad open-weights model with strong instruction following, useful coding and agent capabilities, image and audio input support, and control over how much reasoning it uses.
For most work, start with direct Inkling. It is the faster way to find out whether the model already handles your task well. Move to Inkling Thinking at medium effort when the problem requires planning, comparison, or careful checking. Reserve high and xhigh for tasks that actually improve when the model spends more time and tokens.
As of July 17, 2026, both Inkling variants are pay-as-you-go on NanoGPT and are not included in subscriptions.
Try Inkling or Inkling Thinking on NanoGPT.
