On July 15, 2026, Thinking Machines Lab released its first in-house general-purpose model, "Inkling," as an open-weight model. Until now, the company had offered Tinker, a tool for fine-tuning external models. With its own foundation model now in place, the shape of the business is changing. Instead of renting a finished AI, users can now start from Inkling, train it on their own decision criteria, and take the resulting checkpoints with them.
However, whether the released model can be widely used is a separate matter. Inkling is a massive model with 975 billion total parameters, and even the quantized version requires more than 600GB of combined GPU memory. While the license field says Apache 2.0, a separate terms of use restricts its usage. Thinking Machines' vision of "AI that differs by organization" only comes to life once technical specifications, contracts, and computing resources are all aligned.
A 975-Billion-Parameter MoE, Trained on 45 Trillion Tokens and Over 30 Million RL Iterations
Inkling is a 66-layer decoder-only Transformer with 975 billion total parameters, of which 41 billion are active per token. Its Mixture-of-Experts (MoE) layers contain 256 routed experts and 2 shared experts, routing each token to 6 of the routed experts. Rather than running the entire model every time, this design balances massive scale with controlled computational cost.
The attention mechanism also incorporates design choices aimed at handling long text. Sliding-window layers, which attend to a local range, and global layers, which attend to the entire sequence, are interleaved at a ratio of 5 to 1, supporting a context of up to 1 million tokens. For positional information, the model uses relative position embeddings instead of the now widely used RoPE. Thinking Machines explains that this approach performs better at extrapolating to long inputs.
Images are divided into 40x40 pixel patches and processed through a 4-layer hierarchical MLP. Audio is converted into dMel spectrograms. Both are passed through their own lightweight embedding layers and then processed by the same decoder used for text. The released model accepts text, image, and audio as input, and outputs text. Training used 45 trillion tokens, including video.
The scale of post-training is also substantial. Initial supervised fine-tuning used synthetic data generated by open-weight models such as Kimi K2.5, followed by more than 30 million rounds of reinforcement learning conducted in both synthetic and human-created environments. The aggregated reasoning evaluation score reportedly rose from 0.264 immediately after supervised fine-tuning to 0.356 in the released version. The official announcement describes Inkling as the company's first major training project, and it was trained on NVIDIA GB300 NVL72 systems.
Not the Top Performer, But Adjustable Reasoning Effort
Thinking Machines itself acknowledges that Inkling does not rank first overall among either open or closed models. Instead, it allows the amount of tokens used for reasoning to be adjusted from 0.2 to 0.99, letting users choose accuracy, latency, and cost according to their use case. On Terminal Bench 2.1, the company states it reduced the number of generated tokens needed to reach parity with Nemotron 3 Ultra to about one-third.
The published key evaluation figures make this intent clear.
| Evaluation | Inkling | Comparison Examples |
|---|---|---|
| Humanity's Last Exam (text-only) | 29.7% | GLM 5.2 scores 40.1%, Claude Fable 5 scores 53.3% |
| SWE-Bench Verified | 77.6% | Kimi K2.6 scores 80.2%, DeepSeek V4 Pro scores 80.6% |
| IFBench | 79.8% | GLM 5.2 scores 73.3%, Claude Fable 5 scores 63.5% |
| VoiceBench | 91.4% | Gemini 3.1 Pro scores 94.3% |
On difficult knowledge questions and coding, Inkling falls short of top-tier models, but it is competitive in instruction-following and voice tasks. That said, reading this table as a simple ranking would be misleading. Inkling's scores were measured with the reasoning-effort setting at 0.99 and temperature at 1.0, and the coding evaluation allowed up to 256,000 tokens for the reasoning process. The comparison figures mix results from third-party organizations with self-reported figures from each company, and SWE-Bench Verified used a bash-only harness while Terminal Bench 2.1 used an in-house harness.
Inkling-Small, a preview version unveiled at the same time, clearly reflects this design philosophy. With 276 billion total parameters and 12 billion active at runtime, it scored 46.6% on Humanity's Last Exam with tool use, surpassing the larger model's 46.0%. It also scored 83.4% on IFBench versus 79.8% for the larger model. On the other hand, its SimpleQA Verified score was only 20.9%, a significant drop from the larger model's 43.9%. Scaling down does not uniformly weaken performance, but the strengths and weaknesses shift. Note that the Small version is a preview under testing, and no release date has been set for its parameters.
Even Open-Weight, the Barrier to Running It Locally Is High
Inkling's model card and Hugging Face page list Apache 2.0 as the license, and both a BF16 version and an NVFP4 version for NVIDIA Blackwell are available for download. However, running the BF16 version in-house requires more than 2TB of combined GPU memory. The official reference configurations call for either 8 NVIDIA B300 units or 16 H200 units. Even the NVFP4 version requires more than 600GB, with a W4A4 configuration using 4 B300 units or a W4A16 configuration using 8 H200 units suggested. This is not a scale suitable for testing on a personal workstation.
Contractual terms also require attention. Thinking Machines has established a separate "Model Acceptable Use Policy" that applies to the model's parameters, related materials, and modified versions, and treats acquisition or use of the model as agreement to it. This policy prohibits surveillance, deception, and fully automated decision-making that affects individuals' rights, among other things. The company can update the policy, and continued use is treated as agreement to the updated terms.
These conditions do not align with the Open Source AI Definition 1.0 set by the Open Source Initiative (OSI), which requires that a model can be "used, studied, modified, and shared for any purpose." OSI also requires detailed information about training data and complete code used for training and running the model. The training data documentation released by Thinking Machines is a general description stating that publicly available information, third-party-acquired data, and synthetic data were used, but it does not include an Inkling-specific data inventory or the complete training procedure. Therefore, it is more accurate to call Inkling an open-weight model rather than an open-source AI.
Tinker Bridges the Open Model and Enterprise Data
Given the scale of the execution environment, for many developers the realistic entry point will be Thinking Machines' Tinker or partnering inference services. On Tinker, users can choose between a 64,000-token version and a 256,000-token version of Inkling and fine-tune it using LoRA. A 50% discount is currently available for a limited time. Since Tinker handles the infrastructure operations for distributed training, users can focus on designing their data, evaluation methods, and reinforcement learning environments. The resulting checkpoints can be downloaded, and it is explicitly stated that user data will not be used to train Thinking Machines' own models.
Before Inkling's release, in June, the company demonstrated an early example of this business model through joint research with Bridgewater AIA Labs. Across six financial document classification tasks, the best-performing general-purpose model—tuned by experts adjusting instructions and classification methods—achieved an average score of 78.2%. Using investor evaluation data, they fine-tuned Qwen3-235B via Tinker, raising the average score to 84.7%. Errors reportedly decreased by 29.8%, and the inference cost per task dropped to 1/13.8th of the original.
The foundation model used in this experiment was not Inkling. Even so, it illustrates why Thinking Machines places enterprise-specific evaluation criteria at the center of its business rather than general-purpose rankings. By adding Inkling as a standard model on Tinker, the company has unified foundation model development, a fine-tuning API, checkpoint distribution, and inference partnerships into a single pipeline. Revenue opportunities extend beyond model usage fees to the training workloads generated with each round of iterative fine-tuning.
A Concentrated 1-Gigawatt Investment Supporting Distributed AI
The vision of tailoring AI to each company does not mean shrinking the computing infrastructure involved. In March, Thinking Machines announced a multi-year agreement to deploy at least 1 gigawatt of NVIDIA's next-generation Vera Rubin systems. The target launch is early 2027, and NVIDIA has also invested in the company. Having trained Inkling on GB300 NVL72 systems, the company states it will further expand the computational scale of pre-training, post-training, and reinforcement learning for future models.
What the company is trying to distribute is not the infrastructure for building massive models from scratch, but rather the authority to fine-tune a completed foundation model to match an organization's knowledge and judgment. Inkling is the first product to incorporate the company's own model into this vision. To gauge its success, it will be necessary to track the early-2027 deployment of Vera Rubin and cases where enterprise-specific checkpoints created via Tinker move into real-world operation.