Thinking Machines Lab launched Inkling 2 Small on November 5, 2026, an open-weight AI model that the independent benchmarking site Artificial Analysis says matches Google DeepMind’s Gemini 3.8 Flash (high) on its Intelligence Index. The model is a mixture-of-experts system with 594 billion total parameters and 17 billion active parameters, and Thinking Machines Lab says it plans to release the full weights under an Apache 2.0 license next week, after safety testing finishes.
Table of Contents
Key Takeaways
- Inkling 2 Small is Thinking Machines Lab’s second model generation, launched November 5, 2026.
- It is a mixture-of-experts (MoE) model with 594 billion total parameters and 17 billion active parameters, and it accepts image inputs.
- Artificial Analysis says it scores 31 on its Intelligence Index, matching Gemini 3.8 Flash (high) and surpassing Xiaomi’s MiMo-V2.6-Pro.
- The model is available now through Thinking Machines Lab’s Tinker platform with closed weights.
- Thinking Machines Lab says it intends to release the full weights under an Apache 2.0 license next week, pending completion of external safety testing.
- If the weights ship as planned, Artificial Analysis says it would become the highest-scoring open-weights model developed outside China.
What Is Inkling 2 Small?
Inkling 2 Small is an AI language model built by Thinking Machines Lab, a research company. It is the second model generation the company has released, following its original Inkling Small model. Thinking Machines Lab describes it as a mixture-of-experts model, a design explained below, with 594 billion total parameters and 17 billion active parameters per request, and it supports image inputs in addition to text.
A parameter is one of the internal numerical values a model adjusts during training to learn patterns in data. More parameters generally let a model represent more complex relationships, though how they are organized matters as much as the raw count.
How Does the Mixture-of-Experts Design Work?
A mixture-of-experts (MoE) model splits its parameters into many smaller sub-networks, called experts, and activates only a handful of them for any given input instead of running the entire model every time. This is why Inkling 2 Small has 594 billion total parameters but only 17 billion of them are “active,” meaning used, to process a typical request. The approach lets a model carry a very large total capacity while keeping the computing cost of each answer closer to that of a much smaller model.
How Does Inkling 2 Small Compare to Other AI Models?
According to Thinking Machines Lab, Inkling 2 Small scores 31 on the Artificial Analysis Intelligence Index, a composite benchmark that Artificial Analysis uses to rank AI models across multiple tests. Artificial Analysis, which ran its own independent evaluation, reports that this score matches Google DeepMind’s Gemini 3.8 Flash (high) and surpasses Xiaomi’s MiMo-V2.6-Pro on the same index.
Performance varies by task. Artificial Analysis found Inkling 2 Small scoring above Gemini 3.8 Flash on its AA-3D and Code2Video-AA evaluations, which test 3D-reasoning and video-related coding tasks, while trailing Gemini 3.8 Flash (high) on CritPt v2, a physics and reasoning benchmark, and on AA-Omniscience, which measures factual knowledge and how often a model states incorrect information with confidence. On that hallucination measure, Artificial Analysis recorded a 27% hallucination rate for Inkling 2 Small, down from 61% for the original Inkling Small.
| Benchmark | Inkling 2 Small | Comparison | Source |
|---|---|---|---|
| Artificial Analysis Intelligence Index | 31 | Matches Gemini 3.8 Flash (high) | Thinking Machines Lab, Artificial Analysis |
| AA-Omniscience hallucination rate | 27% | Down from 61% (Inkling Small) | Artificial Analysis |
| AutomationBench-AA | 46% | Up from 5% (Inkling Small) | Artificial Analysis |
| CritPt v2 (physics reasoning) | 56% | Behind Gemini 3.8 Flash high (65%) | Artificial Analysis |
These figures come from two different evaluators using different test sets, so exact numbers should not be averaged across sources. They are useful mainly as a directional picture: substantial gains over the model’s own predecessor, and rough parity with one leading closed model on an overall index, with task-by-task variation.
When Will the Model’s Weights Be Available?
As of launch, Inkling 2 Small is accessible only as a hosted, closed-weights service through Thinking Machines Lab’s Tinker and Tinker Playground tools, which let developers sample outputs and fine-tune the model without downloading it. Thinking Machines Lab states that it intends to publish the full model weights under an Apache 2.0 license, an open-source license that allows commercial use, modification, and redistribution, “next week after the completion of external safety testing and red-teaming.” Artificial Analysis describes this release date as unconfirmed, so readers should treat the timing as Thinking Machines Lab’s stated plan rather than a guaranteed date.
Red-teaming is a security practice in which specialists deliberately try to make a system misbehave, to find weaknesses before the public can exploit them. Thinking Machines Lab says the weights release is pending the results of this testing.
Why Does an Open-Weight Release Matter?
Most leading AI models, including Gemini, Claude, and GPT-series models, are closed-weight: developers can call them through an API but cannot download or inspect the underlying model. An open-weight model can be downloaded, run on a company’s own hardware, modified, and redistributed under its license terms. This matters for organizations with data privacy requirements, researchers studying model behavior, and developers who want to avoid per-token API costs after the initial download.
If Thinking Machines Lab’s weights release goes ahead as described, Artificial Analysis says Inkling 2 Small would become the highest-scoring open-weights model developed outside China on its Intelligence Index. That claim depends on the weights actually shipping and on how competing labs’ models perform at the time, so it should be read as a benchmark snapshot rather than a permanent ranking. Readers who want the fundamentals of how large language models are evaluated and compared can see Future AI Wire’s explainer on why AI hallucination happens and how it is measured.
What Can Developers Do With It Today?
Developers can access Inkling 2 Small now through Tinker for inference and fine-tuning, even before any weights are released. Artificial Analysis lists launch pricing of $0.30 per million input tokens and $1.00 per million output tokens, with cached input priced at $0.03 per million tokens. The model supports text and image inputs, which makes it usable for tasks such as document analysis, coding assistance, and visual question answering through the Tinker interface. Readers building their own automation on top of such models may find Future AI Wire’s guide to AI coding agents useful for understanding how models like this slot into developer workflows, or the explainer on retrieval-augmented generation for how models are connected to external data.
Key Terms
- Mixture-of-experts (MoE): A model architecture that splits parameters into sub-networks and activates only a subset per request, reducing compute cost relative to total model size.
- Open-weight model: An AI model whose trained parameters can be downloaded, inspected, and run independently, under the terms of its license.
- Apache 2.0 license: An open-source software license that permits commercial use, modification, and redistribution with minimal restrictions.
- Intelligence Index: A composite benchmark score published by Artificial Analysis that aggregates results across multiple AI evaluation tests.
- Hallucination rate: The share of a model’s confident, factual-sounding answers that are actually incorrect, as measured on a specific benchmark.
Frequently Asked Questions
What is Inkling 2 Small?
Inkling 2 Small is a mixture-of-experts AI model from Thinking Machines Lab, launched November 5, 2026, with 594 billion total parameters, 17 billion active parameters, and support for image inputs.
Is Inkling 2 Small open source?
Not yet. It launched with closed weights available only through Thinking Machines Lab’s Tinker platform. The company says it plans to release the full weights under an Apache 2.0 license next week, pending safety testing, but Artificial Analysis notes that date is unconfirmed.
How does Inkling 2 Small compare to Gemini 3.8 Flash?
Artificial Analysis says the two models score the same, 31, on its Intelligence Index, though Inkling 2 Small leads on some specific benchmarks, such as AA-3D and Code2Video-AA, and trails on others, including CritPt v2 and AA-Omniscience.
How much does Inkling 2 Small cost to use?
Artificial Analysis lists launch pricing of $0.30 per million input tokens and $1.00 per million output tokens through Tinker, with cached input priced at $0.03 per million tokens.
What does “mixture-of-experts” mean?
It means the model is divided into many smaller sub-networks, called experts, and only a small number of them are used for any single request, which keeps the computing cost per answer much lower than the model’s full size would suggest.
This article reflects reporting available as of November 6, 2026.
Related Reading on Future AI Wire
- AI Hallucination: Why It Happens and How to Reduce It
- What Is RAG (Retrieval-Augmented Generation)?
- AI Coding Agents: What They Are and How They Compare
Sources
- Thinking Machines Lab: Introducing Inkling 2 Small
- Artificial Analysis: Thinking Machines Releases Inkling 2 Small
The Future AI Wire Editorial Team covers artificial intelligence, AI agents, agentic AI, and emerging technologies. We publish clear, research-informed explanations and practical insights to help readers understand AI developments, real-world applications, and the technologies shaping the future.