'Inkling 2 Small Explained: Performance, Weight Size and What to Check Before Release'

Hello, this is Ryuta Hamamoto from TIMEWELL Inc. On November 5, 2026, Thinking Machines Lab announced a new model called Inkling 2 Small¹. The first thing I want to tell you, because it matters most to companies in Japan, is that the weights are not out yet. The announcement says the company intends to release them under Apache 2.0 next week, after external safety testing and red-teaming. As of November 6, 2026, the model is available only as closed weights on a platform called Tinker¹.

In this article I read the official announcement, the model card, third-party rankings and the company's existing models on Hugging Face, and I answer four questions. How much better has it become? How heavy is it? Why do people in Japan ask for open-weight models from outside China? And what should you check first once the weights are released? I am writing for people in IT, procurement and legal, and for executives who set AI policy.

Summary Inkling 2 Small is a mixture-of-experts model with 594B total parameters, of which 17B are used per token, pre-trained natively in NVFP4, a 4-bit floating-point format. It scores 31 on Artificial Analysis Intelligence Index v5.0.1, above Mistral Large 4 at 26, which is the non-Chinese open-weight model we could confirm on the list. The weights are not yet public, and the official pages do not state the precision or file size of the weights to be released. Judging from the company's earlier models, a 4-bit release would likely be in the 300 GB range, which is far beyond a personal computer. Please read this as an article that will be updated after release.

Terms to settle first

There are several technical terms, so here is a table. Come back to it whenever you get lost.

Term	Meaning	Analogy
Open weights	A model whose trained parameters (weights) are published so anyone can run it in their own environment. Different from open source, where code and training data may also be shared	Being handed the finished dish, frozen, rather than the full recipe
MoE (Mixture of Experts)	A structure with many experts, of which only a few are called for each input	A large hospital where only a few specialists see you, depending on your symptoms
Total and active parameters	Total is the number of parts in the whole model. Active is how many are used to process one token	All doctors on staff versus the doctors who actually saw patients that day
NVFP4	NVIDIA's 4-bit floating-point number format. Each number takes 4 bits, so memory drops sharply	Converting a high-resolution photo to a lighter format while keeping it looking the same
Context	How much the model can read at once, counted in tokens (fragments of text)	The size of your desk. A bigger desk lets you spread out more documents
Apache 2.0	A widely used license that broadly permits commercial use and modification, with a patent clause	A permission with a lot of freedom in how you use it
Tinker	Thinking Machines Lab's platform for fine-tuning and inference	The company's own kitchen, where the model is served first

One common misunderstanding about open weights is worth adding. Even when weights are published, the license terms differ from model to model. Some have no restriction on commercial use, and others ask for an extra agreement above a certain scale. For this model, the license on the model card today is Tinker Terms of Use, and Apache 2.0 is still only an intention. Please read the actual text after release.

What was released? Reading the official numbers

Inkling 2 Small is the company's second generation, following Inkling and Inkling-Small¹. According to the model card, it is a 48-layer decoder-only transformer with a mixture-of-experts structure that picks 8 of 384 experts per token. Attention mixes local and global layers. It accepts text and images and outputs text only. There is no audio. The context window reaches 1M tokens, and the image encoder has 292M parameters, based on Perception Encoder¹².

The official description of training is as follows¹.

Stage	Content
H1 (pre-training)	8.4T tokens, 68 hours on GB300 NVL72, trained natively in NVFP4
H2	3.8T tokens, 512K context
H3	About 320B tokens of reasoning and agent trajectories, 1M context
H4	About 34B tokens of high-quality agent trajectories
Reinforcement learning	1,075 steps, more than 17M rollouts, 9 days, 2,048 GPUs

What stands out is the mix of reinforcement-learning tasks. By share of rollouts, coding is 38.5%, tool use and general agents 17.1%, vision 13.9%, and office and knowledge work 9.1%¹. In other words, the emphasis of training was on finishing work with tools, not on being a good conversation partner.

One more statement deserves a careful read. The company says the roughly 34B tokens of H4 were mostly generated by earlier models, including Kimi K3 and Inkling models, then filtered for correctness and alignment¹. When you explain a model's provenance to a business partner, you need to know accurately, within what is disclosed, that the outputs of several models may have been used in training. I see this not as a good or bad point but as a fact showing that explaining provenance rarely fits in one line.

Here are representative results from the official page. All are the developer's own evaluations¹.

Area	Metric	Result
Overall	Vals Index	56.1%
Coding	DeepSWE v1.1	72.1%
Coding	SWEBench Pro v2 (Hard)	86.3%
Agents	MCP Atlas (Full)	85.4%
Vision	MMMU Pro	81.6%
Vision	ScreenSpot Pro	83.8% (91.5% with Python)
Reasoning	HLE-Diamond	28.7%
Long context	AA-LCR v1.1	83.3%
Long context	LongBench v2	70.6%
Factuality	AA-Omniscience non-hallucination rate	73.3%

The list looks strong, but the same page also carries low numbers. FrontierCode v2.0 is 25.4%, the accounting task Apex-Accounting v1.1 is 14.5%, and Terminal Bench Science is 5.7%¹. A fair reading is that it is good at code, tool use and reading images, and still weak at hard accounting and long scientific work.

Where does it stand in third-party rankings?

The developer's numbers alone cannot support a decision, so let us look at an independent evaluator. Artificial Analysis is a third-party site that publishes an Intelligence Index for models. When I checked on November 6, 2026, Inkling 2 Small scored 31 on v5.0.1. It ranked 4th of 29 compared models, against a median of 23 for its class³. That is consistent with the 31 on the official page and the 30.6 in its v5.0 table¹.

The same day, the site's list of open-weight models (index v5.0.1) looked like this⁴.

Model	Developer's location	Index
Kimi K3.1 Pro	China	46
DeepSeek V4.1 Flash (Max)	China	33
GLM-5.3 (Max)	China	32
Inkling 2 Small (for reference, shown as non-open on the site)	United States	31
Mistral Large 4 (High)	Europe	26
Qwen3.8 27B (Xhigh)	China	20

The location column is not the site's wording. It is my own grouping based on where each company is generally known to be based. Here is what the table does and does not support. Inkling 2 Small's 31 is above the 26 of Mistral Large 4, the non-Chinese open-weight model I could confirm on the list. So it is fair to say it is expected to have the highest index among non-Chinese open-weight models once the weights are released. It is not accurate to say it beats all Chinese open-weight models. Kimi K3.1 Pro scores 46, and DeepSeek V4.1 Flash and GLM-5.3 score 32 to 33, all above 31. The site also currently shows Inkling 2 Small as having no public weights, and the list covers only what I could check. Please compare again after release.

The third-party commentary that prompted this article also cited a gain of 18 points over the previous model and parity with Gemini 3.8 Flash. As far as I could check, the index versions differ (the previous Inkling Small appears at 40 on a different version), so I could not compute a difference on the same scale. Because I could not confirm it, I do not use it here.

There is one more trait that index numbers hide. According to the same site, the output tokens used to run the index were 154M against a median of 89.8M, and the time per task was 18.5 minutes against a median of 7.0³. The official limitations also say it lacks concision¹. Longer output and more time raise cost under token billing and lengthen waits. Please look at performance and operating cost separately.

How big are the weights? Separating facts from my calculation

The question I hear most is how big it really is. The short answer is that the official pages do not state a file size. The model card says weight size is not stated, and the same for hardware². So I separate facts from my own calculation.

The facts are these. The total is 594B parameters. It was pre-trained natively in NVFP4. The supported numeric formats are MXFP8 and NVFP4, and an NVFP4 version of the KV cache, the area that holds a long context, is also prepared, with a small effect on evaluation reported¹². The precision and checkpoint size of the weights to be released have not been announced.

What follows is not an official number but my own calculation (an estimate). Weight size is basically the number of parameters times the bytes per parameter.

Precision	Per parameter	Calculation	Weights alone
4-bit (such as NVFP4)	about 0.5 byte	594B × 0.5	about 297 GB, possibly several tens of percent more with scale factors and other data
8-bit (such as FP8)	about 1 byte	594B × 1	about 594 GB
BF16	about 2 bytes	594B × 2	about 1.19 TB

The image encoder's 292M parameters are within rounding error at this scale.

Real examples support this. The company's earlier models are on Hugging Face. On the model page's figures, Inkling in NVFP4 has 975B total parameters and about 592 GB of files, and Inkling-Small in NVFP4 has 276B and about 171 GB⁵⁶⁷. Both include auxiliary files that speed up prediction. That works out to a little over 0.6 bytes per parameter. It is higher than the theoretical 0.5, probably because of scale factors and parts not reduced to 4 bits. Applying that ratio to 594B gives roughly 360 GB. This is only an analogy with earlier models. Please confirm the real value from the file listing after release.

The biggest point of confusion for beginners is the active 17B in MoE. Here are the common misunderstandings.

Misunderstanding	Reality
With 17B active, it needs about the same memory as a 17B model	Memory is set by the 594B total. The chosen experts change per input, so all of them must sit in memory or somewhere quickly readable
Fewer active parameters means the whole thing is light	Only the computation per token is light. The storage footprint is not
At 4 bits it fits on a home PC	Even at 4 bits the weights alone are in the 300 GB range, and a longer context adds KV cache. Using the full 1M tokens needs a large area for it

There are three practical implications. First, this is not a scale that runs on a personal computer or a single typical GPU server. Second, a company in Japan that runs it itself would need a server with multiple GPUs, and for most, the realistic choice will be a cloud inference service or a dedicated deployment. Third, I do not state specific GPU counts or configurations, because the official pages give none. Native NVFP4 computation may require a particular GPU generation (NVIDIA's Blackwell generation). That is my guess too, so it needs checking against official documentation. After release, the safe path is to confirm with measurements and the compatibility tables of inference engines.

At this point you may be asking where such a model can run if you need it in Japan. In Can LLMs run in a Japanese region? I set out model availability by cloud and what regulation actually requires. For the settings and the reality of processing location when you use an open-weight model through an inference service, How to use Hermes Agent uses Fireworks as its example.

Why Japanese companies look for open-weight models from outside China

From here on it is interpretation. I expect calls for open-weight models developed outside China to grow in Japan. I will organize the reasons as company requirements. This is not a claim that models from any particular country are inferior or untrustworthy. In fact, as the table above shows, some Chinese models have a higher index than this new model.

These are the kinds of circumstances in which requirements arise.

Type of requirement	Example
Procurement	A business partner or buyer asks in writing about the developer and location of the AI model used, or conditions appear in a tender or specification
Contracts	Customer contracts or confidentiality agreements ask for representations on data handling or the supply chain
Internal governance	The information-security rules or AI policy set conditions for usable models
Ease of explaining provenance	Whether there are documents to explain training data, derivation and licensing history to auditors or partners
Data location	Where inference runs and where logs remain. Some projects require processing in Japan

In short, there are situations where the choice narrows on criteria other than how good the model is. Until now, asking for high performance while meeting such conditions could leave a narrow range of choices. If Inkling 2 Small publishes its weights, that range widens by one. What widens is the range you can choose from, not a verdict on which is right.

One point to keep in mind is that the developer's country alone does not settle a requirement. As noted above, the training of Inkling 2 Small also included outputs from Kimi K3 according to the company¹. So first put into words whether your requirement concerns the developer's location, the explanation of training data and derivation, the place of inference, or something else. If the requirement stays vague and you judge by a model's nationality alone, you may drift from the facts.

I believe the right approach, for a model from any country, is to judge against your own requirements. If you have several requirements, a wider choice is welcome in itself. Companies whose partners impose different conditions are better served by a design that uses several models and can swap them. A swappable design also helps on cost, which I covered in Five ways to cut the cost of AI coding agents.

The stated limits, and what to check first in practice

Let us sort strengths and weaknesses. The limits the company itself lists are a lack of concision, an occasional failure to pick the best tool in codebases with many skills, no audio, weaker optimization for chat (heavy use of slashes and overreaction to mental-health topics), and falling short of the closed GPT-6.1 Sol on alignment evaluations¹. It also says early checkpoints fabricated results, ignored instructions and concealed actions, and that it addressed this in training¹. The model card adds hallucination, imprecise instruction following, weakness in long multi-turn conversations and performance gaps across languages and dialects, and says medical, legal and safety-critical uses need additional fine-tuning, validation and human review².

For companies in Japan, the last item, performance gaps across languages, matters most. The model card describes its languages as English with general multilingual support². How good it is in Japanese cannot be read from the official numbers. You have to measure on your own documents.

So what should you check first after release? I suggest this order.

1. Read the license text in the original. Did it become Apache 2.0, and are there extra terms or restrictions? Also check that the model card no longer says Tinker Terms of Use.
2. Check the precision and file size of the released weights in the file listing. That is also where you can see whether the estimate of the 300 GB range in this article was right.
3. Decide on candidate inference environments. Inference engine support, the required GPU generation and count, and whether cloud inference services exist and where they process. Counts and configurations should be settled from official documents and measurement.
4. Confirm both by contract and by setting that your data is not used for training. If you use an inference service, read the conditions on log retention and data handling.
5. Try small subjects. Summaries of internal documents, routine research or small code fixes that contain no personal or confidential data. Fix the standard for a correct answer first, and compare against your current model on the same subject.
6. Decide how to evaluate. Prepare 10 to 30 real examples from your work and record not only correctness but also output length, time and cost. Because this model tends to produce long output, recording cost and time matters especially.

Ease of fine-tuning is another point of interest. The company says it learns faster than Inkling and Inkling-Small on the same post-training tasks, and that on a financial summarization task it reached Inkling-Small's final performance in 75 training steps¹. That is material for companies considering fine-tuning on their own data, but it too is the developer's own evaluation.

As of this article, the release of the weights is a plan and has not happened. Once it does, I intend to add the size, the license and third-party rankings to this article.

Summary and next step

Let me sum up.

• Inkling 2 Small is a 594B MoE with 17B active, pre-trained natively in NVFP4, and it accepts image input.
• The third-party ranking (index v5.0.1) is 31, above Mistral Large 4 at 26, the non-Chinese open-weight model we could confirm. Some Chinese models have a higher index.
• The weights are planned for next week and are not out as of November 6, 2026. The current license wording is Tinker Terms of Use.
• The weight size is not stated officially. My estimate is the 300 GB range at 4 bits. Memory is set by the total parameters, and it will not run on a personal computer.
• After release, check the license text, the file size, the inference environment and small measurements, in that order.

What you can do today, before touching the new model, is put your requirements into words. Make one table showing which of the developer's location, the explanation of training and derivation, the place of inference and the handling of data your partners actually ask about. Then you will know what to check after release. Models will keep multiplying. Having your own axis of judgment, instead of a model's name, is the preparation that lasts longest.

At ZEROCK, our commitment is to use only models that do not train on customer data. Projects that require processing in Japan are handled case by case in an individual contract. If you are interested, please see the ZEROCK page, or talk with us through an individual consultation.

Footnotes

1. Thinking Machines Lab, "Inkling 2 Small" announcement (published November 5, 2026; checked November 6, 2026) — https://thinkingmachines.ai/news/inkling-2-small/ ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹ ↩¹² ↩¹³ ↩¹⁴ ↩¹⁵ ↩¹⁶

2. Thinking Machines Lab, "Inkling 2 Small Model Card" (checked November 6, 2026) — https://thinkingmachines.ai/model-card/inkling-2-small/ ↩ ↩² ↩³ ↩⁴ ↩⁵

3. Artificial Analysis, "Inkling 2 Small" (Intelligence Index v5.0.1, checked November 6, 2026; third-party aggregation) — https://artificialanalysis.ai/models/inkling-2-small ↩ ↩²

4. Artificial Analysis, list of open-weight models (Intelligence Index v5.0.1, checked November 6, 2026; third-party aggregation, updated daily) — https://artificialanalysis.ai/models/open-source ↩

5. Hugging Face, "thinkingmachines" organization page (checked November 6, 2026) — https://huggingface.co/thinkingmachines ↩

6. Hugging Face, "thinkingmachines/Inkling-NVFP4" model page and file listing (Apache-2.0 shown, 975B total, repository about 592 GB, checked November 6, 2026) — https://huggingface.co/thinkingmachines/Inkling-NVFP4 ↩

7. Hugging Face, "thinkingmachines/Inkling-Small-NVFP4" model page and file listing (276B total, repository about 171 GB, checked November 6, 2026) — https://huggingface.co/thinkingmachines/Inkling-Small-NVFP4 ↩