Inkling AI Model Explained: Specs, Benchmarks, Hardware and Open Weights
Thinking Machines Lab, the artificial intelligence startup founded by former OpenAI CTO Mira Murati, has released Inkling, its first publicly available general-purpose foundation model. Instead of limiting access to a proprietary API, the company has published the model weights for downloading, customization and self-hosted deployment. [1]
The launch is an important addition to the 2026 open-weight AI surge, but it needs to be understood accurately. Inkling is not a small model that ordinary users can run on a laptop. It is a 975-billion-parameter enterprise-scale system whose smallest official checkpoint still requires at least 600 GB of aggregate GPU memory. [2]
What Is Inkling?
Inkling is a general-purpose multimodal artificial intelligence model developed by Thinking Machines Lab. It accepts text, images and audio as inputs and produces text responses. The company identifies agentic systems, coding assistants, chatbots, tool-use applications and retrieval-augmented generation systems among its intended uses. [2]
It is the first general-purpose model release from Thinking Machines, but it is not the company's first public product. Thinking Machines previously introduced Tinker, a managed platform for customizing and fine-tuning AI models. [5]
That distinction matters because Inkling is not being launched primarily as another consumer chatbot. Thinking Machines is positioning it as a foundation that developers and organizations can adapt to their own data, tools, workflows and operational requirements.
Thinking Machines explicitly acknowledges that Inkling is not the strongest model available overall. Its main proposition is not absolute benchmark leadership. It is a combination of open weights, multimodal capabilities, controllable reasoning and customization through Tinker. [1]
Inkling therefore represents a strategic choice for Murati's new company, but it does not prove that Thinking Machines will release every future model openly. The company may still decide whether to publish or restrict later models based on capability, commercial strategy and safety considerations. [6]
Inkling Key Specifications
| Specification | Inkling |
|---|---|
| Developer | Thinking Machines Lab, Inc. |
| Release date | July 15, 2026 |
| Model type | Multimodal autoregressive transformer |
| Architecture | 66-layer decoder-only sparse Mixture-of-Experts transformer |
| Total parameters | 975 billion |
| Active parameters | 41 billion per token |
| Routed experts | 256 experts, with 6 selected per token |
| Shared experts | 2 shared experts active on every token |
| Maximum context | Up to 1 million tokens |
| Input modalities | Text, images and audio |
| Output modality | Text |
| Pretraining volume | 45 trillion tokens |
| Training content | Text, images, audio and video |
| Numerical support | BF16, MXFP8 and NVFP4 |
| Downloadable checkpoints | BF16 and NVFP4 |
| License listed by provider | Apache 2.0, with a separate acceptable-use policy |
The one-million-token context window is designed for long documents, large code repositories, extended tool workflows and retrieval tasks. However, a large context limit does not guarantee perfect recall or reliable reasoning across every included token. Actual performance still depends on prompt structure, retrieval quality, inference settings and the task being performed.
Inkling's released form accepts text, image and audio inputs, but it currently generates text output. Video was included among its pretraining content, although video is not listed as a direct input modality in the released model card. [1] [2]
How Inkling's Architecture Works
Inkling uses a sparse Mixture-of-Experts architecture. The model contains 975 billion parameters in total, but only around 41 billion are active for each token. Instead of processing every token through the entire network, the model routes each token to selected specialist components.
Each Mixture-of-Experts layer contains 256 routed experts and 2 shared experts. Six routed experts are selected for each token, while the shared experts remain active. This approach is intended to provide the capacity of a very large model without activating all of its parameters for every request. [2]
Input processing
Text, image and audio information is converted into representations that can be processed inside a shared hidden space.
Expert routing
Each token is directed to six selected routed experts, together with two shared experts.
Hybrid attention
Local and global attention layers are combined to support efficiency and long-context processing.
Text generation
The decoder combines the processed information and generates a text response.
Training and post-training
Thinking Machines says Inkling was trained from scratch and pretrained on 45 trillion tokens containing text, images, audio and video. The company used NVIDIA GB300 NVL72 systems for the training effort. [1]
The model was then post-trained across mathematics, coding, tool use, audio, image understanding, conversation and safety tasks. Thinking Machines reports that an initial supervised fine-tuning stage used synthetic data generated partly by existing open-weight models, including Kimi K2.5. Most of the later post-training compute was reportedly used for reinforcement learning across synthetic and human-created environments.
Inkling also supports adjustable reasoning effort. Developers can change how much computation and output-token budget the model uses, creating a tradeoff among response quality, latency and operating cost.
Inkling Benchmarks: Strengths and Weaknesses
Thinking Machines published benchmark results covering reasoning, coding, tool use, factuality, instruction following, vision, audio and safety. The results show a capable general-purpose model, but they do not establish Inkling as the overall performance leader.
The results below were published by Thinking Machines. Evaluations were generally run at a reasoning-effort setting of 0.99 and a temperature of 1.0. Some scores came from external evaluators, while others used the company's own harnesses. Independent production testing may produce different results. [1]
| Area | Benchmark | Inkling score | Assessment |
|---|---|---|---|
| Mathematics | AIME 2026 | 97.1% | |
| Advanced reasoning | GPQA Diamond | 87.2% | |
| Coding | SWE-bench Verified | 77.6% | |
| Instruction following | IFBench | 79.8% | |
| Multilingual knowledge | Global-MMLU-Lite | 88.7% | |
| General agents | Tau 3 Banking | 23.7% | |
| Factuality | SimpleQA Verified | 43.9% | |
| Extreme reasoning | Humanity's Last Exam, text only | 29.7% |
Inkling's strongest reported areas include mathematical reasoning, instruction following and competitive software-engineering performance. It also performs well enough across multiple categories to serve as a broad foundation for further customization.
The weaker results are equally important. A 43.9 percent score on SimpleQA Verified indicates that factual question answering remains a material limitation. Its 23.7 percent result on Tau 3 Banking also shows that complex general-agent tasks remain difficult.
This balanced view matches the company's own positioning. Inkling can be competitive without being the strongest model overall. Reuters similarly reported that other models retained the performance advantage across the complete benchmark set, although Inkling showed promise on selected agent-related tasks. [5]
What the benchmark table really means
Inkling should not be selected merely because it reaches a high score on one mathematics or coding benchmark. Its business value will depend on whether customization produces better results for a specific workload at an acceptable cost.
Open-Weight Does Not Mean Completely Open-Source
Inkling is best described as an open-weight model. Its trained parameters are available through Hugging Face, allowing developers to download, inspect, host, fine-tune and integrate the model into their own systems. [3]
Open weights are valuable because they reduce dependence on a single proprietary API provider. Organizations can choose their own infrastructure, customize the model privately and potentially move between deployment providers without rebuilding an entire application around a different closed model.
What open weights provide
- Access to trained model parameters
- Private or self-hosted deployment
- Fine-tuning on specialized data
- More control over hosting and infrastructure
- Integration into third-party applications
What open weights do not guarantee
- Complete disclosure of the training data
- Full reproducibility of the training process
- Low-cost deployment
- Operation on consumer hardware
- Freedom from separate usage conditions
Thinking Machines identifies Apache 2.0 as the model license, but it also publishes a separate Model Acceptable Use Policy covering the model weights, parameters, associated materials and modified versions. [2] [4]
Organizations considering commercial deployment should review both documents and obtain appropriate legal advice when necessary. The existence of an Apache 2.0 label should not be treated as a substitute for examining the additional use conditions.
Can Inkling Run Locally?
Technically, Inkling can be downloaded and run on infrastructure controlled by the user. Practically, the complete model cannot run on an ordinary desktop computer or laptop.
Official example configurations:
- 8 NVIDIA B300 GPUs
- 16 NVIDIA H200 GPUs
Official example configurations:
- 4 NVIDIA B300 GPUs in W4A4 mode
- 8 NVIDIA H200 GPUs in W4A16 mode
"Run locally" does not mean "run on your laptop." In Inkling's case, local deployment generally means operating the model on an enterprise GPU cluster, private cloud environment or data center.
Thinking Machines lists SGLang, vLLM, TokenSpeed, Unsloth and Hugging Face among the supported deployment frameworks. Even with appropriate software, the infrastructure requirement places the full model beyond the reach of most individual developers and small organizations. [2]
Most users are therefore more likely to access Inkling through Tinker or a third-party inference service. Open weights increase deployment choice, but they do not eliminate the economic role of hosted platforms.
Inkling and the Tinker Customization Platform
Inkling is closely connected to Tinker, Thinking Machines' managed model-customization platform. Tinker is intended to handle much of the distributed infrastructure required for fine-tuning while allowing developers to build specialized training and evaluation workflows.
Thinking Machines demonstrated a controlled workflow in which Inkling used Tinker tools to prepare training data, write a fine-tuning job, evaluate the updated checkpoint and then load the customized model. [1]
This demonstration does not mean Inkling independently rewrote itself without supervision. It means the model was given access to a defined fine-tuning environment, tools, instructions and evaluation criteria.
This model-plus-platform strategy may be central to Thinking Machines' business. The open weights encourage adoption, experimentation and independent deployment, while Tinker provides a managed service for organizations that want customization without operating the complete training infrastructure themselves.
Thinking Machines also lists Together AI, Fireworks, Modal, Databricks and Baseten among its inference partners. This shows that the open-weight market can still rely heavily on paid APIs and managed infrastructure. [1]
Does Inkling Threaten Pay-Per-Token AI APIs?
Inkling increases competitive pressure on proprietary AI providers, but it does not make pay-per-token APIs obsolete.
Open-weight systems can weaken vendor lock-in because organizations have more control over customization, deployment and hosting. A company can fine-tune Inkling on specialized data, operate it within controlled infrastructure or select among several hosting providers.
That flexibility may make it more difficult for closed-model providers to maintain premium prices when an open-weight alternative performs adequately for a specific task.
However, operating a model of this scale involves substantial costs. Hardware, cloud capacity, networking, electricity, inference optimization, security, monitoring and specialist engineering can make self-hosting more expensive than an API for low-volume or moderate workloads.
Open-weight deployment
- Greater control over model and data
- More customization options
- Reduced dependence on one model vendor
- Potential savings at sufficient scale
- Higher infrastructure responsibility
Hosted API deployment
- Faster initial implementation
- No private GPU cluster required
- Simpler scaling and maintenance
- Lower operational burden
- Continued dependence on service providers
The more likely market outcome
Inkling may not destroy the API business model. It may shift part of the market from APIs controlled by closed-model developers toward paid services that host, fine-tune and optimize open-weight models.
The AI market is therefore likely to support several deployment models at the same time: premium closed APIs, hosted open-weight inference, privately deployed enterprise systems and smaller models running directly on personal devices.
Inkling adds pressure to proprietary providers, but the competition is not simply open weights versus APIs. The real contest involves intelligence, customization, privacy, operating cost, reliability and ease of deployment.
Safety, Bias and Known Limitations
Thinking Machines reports evaluating Inkling across harmful manipulation, dangerous capabilities, cyber risks, vulnerable-user interactions and loss-of-control scenarios. The company says it used internal evaluations and external safety testers. [1]
These results should still be interpreted cautiously. At the time of publication, most detailed safety and performance evidence comes from Thinking Machines itself. Independent red-team testing and long-term production evidence remain limited because Inkling has only recently been released.
Known practical limitations
- Inkling can produce incorrect or unsupported information.
- Factual question answering remains weaker than that of several leading closed models.
- It may fail to follow complex instructions consistently.
- Performance can decline during long multi-turn conversations.
- Training-data bias may affect languages, cultures and demographic groups unevenly.
- Fine-tuning can alter both capability and safety behavior.
- High-stakes decisions still require qualified human review.
Open weights also transfer more responsibility to downstream developers. Organizations modifying or deploying Inkling must assess whether the customized system remains appropriate for its users, data, industry and operating environment.
The Model Acceptable Use Policy includes restrictions involving illegal activity, child exploitation, violence, harmful cyber activity, privacy violations, discrimination, fraud, misinformation and unauthorized professional practice. [4]
What Is Inkling-Small?
Thinking Machines is also previewing Inkling-Small, a 276-billion-parameter Mixture-of-Experts model with 12 billion active parameters. Its full weights had not yet been released when this article was published. [1]
The smaller model is designed to offer a different balance among performance, cost and latency. According to company testing, Inkling-Small approaches the larger model on several reasoning and coding benchmarks and exceeds it on selected tests, including IFBench.
Inkling-Small may ultimately have broader practical importance than the full model. Twelve billion active parameters could make it more suitable for cost-sensitive coding, evaluation, synthetic-data generation and other workloads where latency matters.
That conclusion remains preliminary until the complete weights, hardware requirements and independent benchmarks are available.
Frequently Asked Questions
What is Thinking Machines Inkling?
Inkling is a 975-billion-parameter multimodal Mixture-of-Experts model developed by Mira Murati's Thinking Machines Lab. It accepts text, image and audio inputs and generates text.
Is Inkling Thinking Machines' first product?
No. Inkling is the company's first general-purpose foundation model. Thinking Machines previously released Tinker, its platform for customizing AI models.
Is Inkling open-source?
Inkling is more precisely described as open-weight. Its trained parameters are downloadable, but the complete training data and full model-development process have not been released in a form that would make the system fully reproducible.
Can Inkling run on a laptop?
No. The BF16 checkpoint requires at least 2 TB of aggregate GPU memory. The quantized NVFP4 checkpoint still requires at least 600 GB of aggregate GPU memory.
How many parameters does Inkling use?
Inkling contains 975 billion parameters in total but activates approximately 41 billion parameters for each token through its sparse Mixture-of-Experts architecture.
Does Inkling support images and audio?
Yes. The released model accepts text, image and audio inputs. Its output is currently text.
Is Inkling better than leading closed models?
Not overall. Inkling performs strongly on selected mathematics, coding and instruction-following tests, but it trails leading models on several advanced reasoning, factuality and agentic benchmarks.
Will Inkling replace paid AI APIs?
Unlikely. Inkling increases deployment choice and pricing pressure, but its hardware requirements mean that many users will still rely on paid hosting, managed inference or Tinker.
What is Inkling-Small?
Inkling-Small is a previewed 276-billion-parameter model with 12 billion active parameters. Thinking Machines says the full weights will be released after testing is completed.
Inkling Is a Serious Open-Weight Release, Not a Laptop AI Model
Inkling does not immediately displace the strongest systems from OpenAI, Anthropic, Google or leading open-weight laboratories. Its factuality and general-agent results expose meaningful weaknesses, while its enormous memory requirements limit true self-hosting to organizations with substantial infrastructure.
It is still a significant release. A heavily funded American AI startup led by one of the industry's most prominent executives has made its first general-purpose foundation model downloadable, customizable and deployable outside a single proprietary platform.
The most important part of Thinking Machines' strategy is not the claim that Inkling is the smartest available model. It is the argument that an organization may obtain more value from a model it can control and specialize than from continually renting a general-purpose model designed by another company.
Inkling will place pressure on closed-model providers to improve pricing, privacy, customization and deployment flexibility. However, it will not end paid AI services. Its scale may instead create more demand for companies that can host and optimize open-weight models.
Inkling is therefore best understood as an opening move. Its long-term importance will depend on whether developers can customize it effectively, whether hosted and private deployments remain economical and whether later models from Thinking Machines improve on its current weaknesses.
Sources and Further Reading
- Thinking Machines Lab - Inkling: Our Open-Weights Model
- Thinking Machines Lab - Inkling Model Card
- Hugging Face - Thinking Machines Inkling Repository
- Thinking Machines Lab - Model Acceptable Use Policy
- Reuters - Thinking Machines Launches an Open-Weight AI Model
- Axios - Mira Murati's Thinking Machines Debuts First AI Model
Editorial note: Benchmark results and technical specifications were checked against information available on July 18, 2026. Most detailed benchmark and safety findings remain provider-reported and may change as independent evaluations become available.