Inkling 2 Small is Thinking Machines Lab's open-weight multimodal model: 594 billion total parameters with 17 billion active per token, text and image input, and an upstream context of up to 1M tokens. It was built for long work that ends in something you can check, such as a financial analysis, a research answer or a working web page. Its maker says it is less optimized for chat; it is a work model more than a chat model.
We offer Inkling 2 Small in Maple Chat, Maple Work and Maple Proxy on paid plans (Pro, Max, Max+ and Team), with a 524,288-token context window. This guide covers what it's good at, where it falls short, what self-hosting takes, and which route fits your work.
| What is Inkling 2 Small? | Thinking Machines Lab's open-weight multimodal model: 594B total parameters, 17B active, up to 1M tokens of context, text and image input.1 |
|---|---|
| Is it open source? | It is open-weight under Apache 2.0,2 with Thinking Machines' Model Acceptable Use Policy on top.3 |
| What is it good at? | Long work tasks: finance and document analysis, research that needs many searches, and design and web builds. |
| What is it weak at? | Casual chat, short answers and terminal-heavy coding. |
| What hardware does it need? | At least 500 GB of combined GPU memory for the main checkpoint, such as two NVIDIA B300 or four H200 GPUs. |
| Is it in Maple? | Yes: Maple Chat, Maple Work and Maple Proxy on paid plans, with a 524,288-token context window. |
| Is it private? | That depends on the route, not the model. |
What is Inkling 2 Small?
It is a mixture-of-experts model: 48 layers, with each token routed to 8 of 384 experts, which is why only 17B of its 594B parameters do the work for any one token. Attention mixes local and global layers, and a vision encoder feeds images into the same model as text.1 It reads text and images and writes text. It doesn't take audio.4
Active parameters describe compute per token, not memory. Any token can route to any expert, so all 594B parameters have to be available in memory, not just the 17B a token uses.
You can set how hard it thinks. Thinking Machines publishes six named effort presets, from none to xhigh, with high as the default.5 Thinking Machines ran its own evaluations at xhigh.4 More effort means longer reasoning, which matters for a model that already writes a lot.
Thinking Machines is a U.S. lab, and Plotly called Inkling 2 Small "the most accurate open-weight model from a U.S. lab we've tested" on its data-analysis exam.6
The Inkling mark on this guide's cover belongs to Thinking Machines Lab and is shown to identify the model. The Inkling family's own page is on their site.7
Is Inkling 2 Small open source?
It is open-weight. The model card lists the license as Apache 2.0 and the repository links to the standard Apache text, though the repository has no license file of its own.8 Apache 2.0 allows use, modification and redistribution, including commercial use.
Thinking Machines first offered the model only through its Tinker platform, then published the weights on Hugging Face.4 The repository holds the NVFP4 checkpoint and links an MXFP8 one.8
The weights come with a second document. Thinking Machines' Model Acceptable Use Policy applies to anyone who touches them, including modified versions:
By accessing, downloading, or using any Model Materials, you agree to be bound by this Model AUP, unless we have otherwise agreed in writing.
Three of its clauses matter for ordinary builders:
- Safeguards stay on. You may not circumvent, disable or remove measures built into the model.
- No unlicensed advice. You may not use it to give unlicensed medical, legal or financial advice, or to automate consequential decisions in those areas without human oversight.
- Disclose its limits. If you put it in a product, you agree to tell users about its known material limitations.
If you plan to build on the weights, read both documents and get legal advice where it matters.
What is Inkling 2 Small good at?
Long tasks that end in a deliverable, and finance most of all. Thinking Machines puts it plainly:
Inkling 2 Small is especially strong in finance and quantitative reasoning over spreadsheets, filings and documents.
Its model card compares it with 12 other models, and it leads on about 20 rows. Here are five, all Thinking Machines' own results at xhigh effort.1, 4
| Benchmark (Thinking Machines-reported) | Inkling 2 Small | Next best on the card | What it measures |
|---|---|---|---|
| Finance Agent v2.1 | 67.5% | 65.8% (Gemini 3.8 Flash) | Financial research tasks with tools |
| BrowseComp (Inkling with compaction) | 90.2% | 89.4% (Gemini 3.8 Flash) | Finding hard-to-locate facts on the web |
| AA-Omniscience non-hallucination rate | 73.3% | 72.0% (Muse Spark Open) | How often it holds back instead of answering wrongly |
| DesignArena 3D Design | 1457 | 1367 (Kimi K3) | People's votes on 3D builds |
| DesignArena Game Dev | 1424 | 1351 (Kimi K3) | People's votes on browser games |
Read the table with three caveats. The comparison set is open-weight Flash and Pro models, up to Kimi K3, plus three small closed ones (Claude Haiku 5.5, GPT-6 Luna and Gemini 3.8 Flash), with no frontier closed models and no Kimi K3.1 Pro. The BrowseComp row labels compaction, which condenses earlier context so a long search can continue, for Inkling's score only. And Artificial Analysis found that Inkling 2 Small "hallucinates less mainly because it answers fewer questions, while accuracy improves only slightly."9
Independent tests point the same way on finance. Vals AI ranks it fifth of 41 models on Finance Agent v2.1 at 67.54%, and its launch evaluation called it the top open-weight model on that test.10 Plotly's exam of real data-analysis questions gave it 39 of 43, with a median of 13 seconds a question.6
Where is Inkling 2 Small weak?
In conversation, and Thinking Machines says so in its own limitations section:
Inkling 2 Small is also less optimized for the chat experience. We notice issues such as worse organization and presentation of content, overuses of slashes, and overreaction to mental health and other safety related queries.
The same section adds that it "is not the most succinct communicator." Independent measures agree, and the overall rankings put it mid-pack rather than at the top:
| Signal | Inkling 2 Small | Context |
|---|---|---|
| Artificial Analysis Intelligence Index | 31 on Artificial Analysis' page11 | The card's own row (v5.0) shows 30.6, behind Claude Haiku 5.5 (36.5), DeepSeek V4.1 Flash (32.6), GLM 5.3 (31.7) and Gemini 3.8 Flash (30.8)1 |
| Vals Index | 56.08%, 23rd of 3810 | Vals' overall index across its benchmarks |
| Terminal-Bench 4.0 | 13.1% on the card;1 12.63%, 45th of 65 on Vals10 | GLM 5.3 scores 41.9% on the card |
| Speed and length | 21.6 minutes per index task; 154M output tokens across Artificial Analysis' full index run11 | Its medians for similar open models: 13.7 minutes and 104M |
| FORTRESS (benign) | 79.7%, the lowest on the card1 | The other 12 models score 84.0% to 97.4% |
The FORTRESS row needs a word. Its benign set checks that a model still answers harmless prompts, so a lower score means more harmless requests refused. Combined with the verbosity and the roughly 38-second wait before the first word of an answer that Artificial Analysis measured,11 that is why a quick question can feel slow and overcautious.
That makes it a specialist, and a low-cost one: at $0.45 per million input tokens, feeding it long filings and documents stays affordable. It pays to choose it for the right job.
Which tasks fit Inkling 2 Small?
The evidence above sorts most work into a clear yes or no:
| Task | Fit | Why |
|---|---|---|
| Financial analysis over filings and spreadsheets | Good | Leads Finance Agent on its card; top open-weight model on Vals' Finance Agent at launch |
| Research that needs many searches | Good | Leads BrowseComp on its card, with compaction |
| Reading documents, charts and screenshots | Good | 94.6% on OmniDocBench and 83.8% on ScreenSpot Pro (card) |
| Web pages, games and 3D builds | Good | Leads all four DesignArena rows on its card |
| Quick questions and casual chat | Poor | Its maker's own limitation; slow to start and verbose |
| Emotional conversations | Poor | Its maker notes overreaction to mental health queries |
| Terminal-heavy coding agents | Poor | 13.1% on Terminal-Bench 4.0 |
To get the most from it, hand it the whole job at once: the files, the question, and the form you want the answer in, such as a table, a memo or a page. Expect a long answer and check its numbers before you rely on them, as the model card itself advises for high-stakes use.
For quick everyday questions, a model tuned for chat such as DeepSeek V4.1 Flash is a better fit. For hard coding and agent work in a terminal, Kimi K3.1 Pro is the stronger choice. Our model guide covers how to pick a model and thinking level for a task.
What hardware does Inkling 2 Small need to self-host?
The model card gives a clear floor. The main NVFP4 checkpoint "requires a GPU cluster with at least 500GB of aggregated VRAM," such as two NVIDIA B300 or four H200 GPUs. The alternative MXFP8 checkpoint needs at least 800 GB, such as four B300 or eight H200 GPUs.1 Thinking Machines trained the model in NVFP4 and recommends that checkpoint for every use.
We added up the files in the NVFP4 repository: 24 weight files totaling about 390 GB, plus a 3.4 GB multi-token prediction file.8 The rest of the memory budget goes to the context you serve and how many requests run at once.
Context is where the hardware grows. vLLM's recipe verified one four-GPU NVIDIA GB200 tray for contexts up to 262,144 tokens, and two trays for the full 1,048,576.12
Community quantizations are smaller. Unsloth's smallest GGUF needs 171.7 GB of disk for the weights alone, and running it in llama.cpp relies on support that was still a pending upstream change. Unsloth notes that file size is not the total memory requirement.13 Treat these as experiments, not as the maker's supported path, and test quality on your own work.
Self-hosting is the right call in some cases:
- your information must stay on hardware you control
- you want to fine-tune or modify the model
- you already run multi-GPU inference with capacity to spare
If none of those apply, the operational work may not buy you anything a hosted route can't.
How do you use Inkling 2 Small in Maple?
Choose Inkling 2 Small from the model picker in Maple Chat or Maple Work, our desktop app, on a paid plan: Pro, Max, Max+ or Team. It isn't on the Free plan. Plan details are on our pricing page, and the model sits with the others we offer in our model library.
Maple Work is the natural home for it. Long, multi-step tasks with files are what Inkling 2 Small does best, and Work is built for that kind of job. Use Maple Chat when you want to ask about one document or image without setting up a task. Work lists it under Fast & cheap: it costs little per token, but it writes long answers, so a whole task can take longer than with a model built for chat.
Reasoning is always on, and you can pick one of five effort levels: minimal, low, medium, high (the default) or xhigh. We leave out "none" because it doesn't reliably turn reasoning off; on hard prompts, the model starts writing its working into the answer instead.14 Lower effort gives shorter reasoning, which can speed up simpler jobs.
You can upload an image while using Inkling 2 Small in Maple Chat. The app first has a separate vision model describe it, then gives that description to Inkling 2 Small. The answer is based on the description, not a direct view of the image, which matters when a small visual detail decides the task.15 Through Maple Proxy, our documentation lists text and image input for Inkling 2 Small.
For your own pipelines, Maple Proxy is our API route, with the model ID inkling-2-small. It's billed per token at these rates:16
| Model | Input | Cached input | Output |
|---|---|---|---|
Inkling 2 Smallinkling-2-small | $0.45 | $0.05 | $1.50 |
USD equivalent per 1 million tokens. API usage is charged by the chosen model’s input and output tokens. It draws from your plan’s included usage first, then from credits you add. Cached-input rates apply only to tokens reported as cached. — means the standard input price applies. Automatic model aliases are billed at the rate of the model selected for the request.
Check the Proxy documentation for current input types and limits.
Plan for three limits in Maple:
- Context is 524,288 tokens, half the model's 1M upstream maximum.
- Reasoning can't be turned off. Minimal is the lowest effort.
- In Chat, images arrive as a description from a separate vision model.
This route gives you two things. First, you use a 594B model without buying, running or updating multi-GPU servers. Second, your requests go through our protected inference path, not a conventional hosted API; the next section explains it.
What information reaches each route?
That depends on who operates the model: you, a conventional host, or Maple. Several questions hide inside "is it private": who can read your prompts while they're processed, what happens to them afterward, and whether any of it trains a model.
Self-hosted weights. Prompts and outputs can stay on infrastructure you operate, as long as your tools and connected services don't send them elsewhere. The trust boundary is yours, along with the job of securing it.
A conventional hosted API. With Thinking Machines' Tinker platform or another host, your requests are processed in a form the provider's systems can read. Retention, training use and staff access then depend on that provider's terms, which vary by service and plan.
Maple. Your prompts are encrypted before they leave your device and processed inside a secure enclave, an isolated computing environment. The model runs in its own attested enclave, and Maple's enclave verifies it before forwarding a prompt, much as certificate chains work on the web. Our proof page describes the architecture and the attestation you can check yourself.

That design protects the contents of your chats: your prompts and the responses. Our privacy notice covers everything outside them. It also doesn't vouch for what the model's output leads you to do. If an agent runs code Inkling 2 Small wrote, or you act on its financial analysis, review that work as you would any untrusted input.
Which route fits your task?
It depends on where your information may go, how much context you need and how much infrastructure you can run.
| Factor | Self-hosted weights | Tinker or another host | Maple |
|---|---|---|---|
| Setup | At least 500 GB of GPU memory; build and run serving | Account and API key | Any paid plan; choose Inkling 2 Small |
| Context | Up to 1,048,576 tokens, subject to your hardware | Per provider | 524,288 tokens |
| Input | Text and image | Per provider | Text and image; in Chat, images reach it as a description |
| Thinking effort | None to xhigh | Per provider | Minimal to xhigh; reasoning always on |
| Data boundary | Your infrastructure | Provider's servers, under its terms | Encrypted on your device, decrypted only inside a hardware-isolated enclave |
| Cost source | Hardware, power and engineering time | Provider's pricing page | One Maple balance for Chat, Work and the API, at per-token prices |
| Main limitation | Operating cost; the Model Acceptable Use Policy applies | Provider can read requests; terms vary | Half the upstream context; paid plan required |
A rough decision rule:
- Choose self-hosting if information must stay on your hardware or you want to fine-tune the weights.
- Choose a host that serves the full 1M window if you need it and a conventional provider boundary fits your data.
- Choose Maple if you want Inkling 2 Small on sensitive work, such as client financials or internal documents, without running infrastructure.
Whichever route you pick, test it on the work you'd hand it, not on small talk. That is where this model earns its place.
Sources
- Thinking Machines Lab, Inkling 2 Small model card. Parameters, architecture, modalities, context, license, hardware requirements and evaluation results against the comparison set.
- Apache Software Foundation, Apache License, Version 2.0, January 2004. The license text the model's repository links to.
- Thinking Machines Lab, Model Acceptable Use Policy, July 15, 2026. Conditions on using the weights, including safeguards, professional advice and disclosure.
- Thinking Machines Lab, Introducing Inkling 2 Small, November 5, 2026. Launch post: capabilities, evaluation settings, the weights release and the model's limitations.
- Thinking Machines Lab, Thinking effort, Tinker cookbook. Inkling's named effort presets and the default.
- Plotly, Inkling 2 Small Benchmark Results for Data Analysis, November 12, 2026. Data-analysis exam results and the comparison with other open-weight models.
- Thinking Machines Lab, Inkling. The Inkling model family and its mark.
- Thinking Machines Lab, Inkling-2-Small repository on Hugging Face. NVFP4 checkpoint files, chat template effort levels and the link to the MXFP8 checkpoint.
- Artificial Analysis, Thinking Machines Inkling 2 Small: the highest scoring open weights model outside of China pending weights release, November 5, 2026. Index result, hallucination analysis and output-token use.
- Vals AI, Inkling 2 Small benchmarks. Vals Index, Finance Agent v2.1 and Terminal-Bench 4.0 rankings, and the launch evaluation.
- Artificial Analysis, Inkling 2 Small: Intelligence, Performance and Price Analysis. Intelligence Index score, time per task, time to first answer token and verbosity.
- vLLM, Inkling-2-Small recipe. Verified GB200 layouts for 262,144-token and 1,048,576-token contexts.
- Unsloth, Inkling 2 Small: how to run locally. GGUF quantization sizes, llama.cpp support and memory notes.
- Maple, public model configuration. Plan tier, context window, effort levels, sampling and inputs for Inkling 2 Small in Maple.
- Maple, Maple Chat image handling and vision helper. Current-turn images are described by a separate vision model before the chosen model runs.
- Maple, Pricing. Per-token rates for Inkling 2 Small and the plans that include it.


