Thinking Machines Lab officially released its multimodal AI "Inkling-Small" as an open-weight model on July 30, 2026. When the large model "Inkling" was announced on July 15, only a preview version was available, and no release date for the weights had been set. This time, the company published all weights on Hugging Face and enabled fine-tuning via Tinker as well as use through the Playground. Notably, the model—shrunk to under 30% of the large version's size—actually outperformed the large model on some reasoning and coding evaluations.
However, "Small" is a relative label within the company's product lineup. Even the quantized version requires a total of 180GB or more of GPU memory, meaning this is not a model that runs on an ordinary PC. Benchmarks also show that its knowledge base and factuality fall short of the large version. Inkling-Small represents a more pragmatic design decision about which capabilities to preserve in order to lower the operating costs of enterprise AI.
Two Additional Weeks of Training After Preview, Now With Full Weights Released
Inkling-Small has a total of 276 billion parameters, with 12 billion active parameters used per token processed. The large Inkling, by comparison, has 975 billion and 41 billion respectively. Thinking Machines describes it as "a quarter of the size," but the official figures put the ratio at 28.3% for total parameters and 29.3% for active parameters. Strictly speaking, it's more accurate to say roughly 30% of the large version.
Despite the smaller size, the core architecture remains the same. It's a 42-layer decoder-only Transformer, where the Mixture-of-Experts (MoE) layers route each token to 6 out of 256 selectable experts, while always using 2 shared experts. In addition to text, it can accept image and audio input, and supports a context window of up to 1 million tokens. Images are divided into 40×40 pixel patches and processed through a 4-layer hMLP, while audio is converted into dMel spectrograms.
The reason performance was preserved isn't simply that an existing model was trimmed down as-is. Because training for Inkling-Small began after the large version, the company was able to refine the pre-training data mix and training procedures. Post-training of the preview version also partly used on-policy distillation, with the large Inkling serving as the teacher. Furthermore, starting from the preview version, the team continued reinforcement learning for two weeks to strengthen the model's ability to write code by operating tools. The official release is a separate checkpoint taken after this additional training was completed.
Areas Where the ~30%-Sized Model Outperformed the Large Version
According to the company's own comparisons, Inkling-Small outperformed the large version on several reasoning and coding metrics. The main differences are as follows:
| Evaluation | Inkling-Small | Inkling |
|---|---|---|
| SWE-bench Verified | 80.2% | 77.6% |
| Terminal Bench 2.1 | 64.7% | 63.8% |
| HLE (no tools) | 31.6% | 29.7% |
| IFBench | 82.2% | 79.8% |
| SimpleQA Verified | 20.6% | 43.9% |
| Tau 3 Banking | 15.5% | 23.7% |
These results show that the effects of downsizing are not uniform. Small outperforms on SWE-bench Verified, which measures the ability to solve software engineering tasks, and on IFBench, which measures instruction-following. On the other hand, SimpleQA Verified—which deals with short factual questions—dropped to less than half. Thinking Machines itself acknowledges that the large Inkling has the edge in knowledge coverage and factuality. The choice of model should differ depending on whether the use case combines retrieval or internal company data, versus relying directly on the model's internal knowledge.
The comparison conditions also involved certain constraints. Both models were set to an effort value of 0.99 representing reasoning intensity, with temperature set to 1.0, and coding evaluations allowed up to 256,000 tokens of processing. SWE-bench Verified used a bash-only harness, while Terminal Bench 2.1 used an internal harness. Therefore, the numbers in the table shouldn't be taken as definitive proof of superiority across all real-world tasks.
At Artificial Analysis, an independent evaluator, the Intelligence Index v4.1—which aggregates nine different evaluations—scored 40 for Inkling-Small and 41 for the large Inkling. The fact that an overall gap of just one point was observed even in measurements independent of the vendor lends support to the claim that the smaller model preserves performance. That said, even when the aggregate scores are close, significant weaknesses like the one seen in SimpleQA Verified don't disappear.
The Self-Hosting Barrier, Now Lowered to 180GB
The requirements for self-hosting have dropped significantly compared to the large version. The BF16 version requires a total of 600GB or more of GPU memory, with the official configuration calling for either 4 NVIDIA B300 units or 8 H200 units. With NVFP4 quantization, this drops to 180GB or more, using 1 B300 for W4A4 or 2 H200 units for W4A16. Since the large Inkling requires 2TB or more in BF16 and 600GB or more even with NVFP4, Small's requirements come out to roughly 30% of the large version's in both cases.
Even so, this remains far removed from a personal laptop or a typical workstation. Even though MoE limits the active parameters used per run to 12 billion, the full 276 billion parameters must still be held in GPU memory. In practice, the realistic options are limited to enterprise GPU servers, third-party inference providers, or Thinking Machines' own Tinker.
Tinker's Serverless Inference is priced at $0.30 per million input tokens and $1.20 per million output tokens. Since the large Inkling costs $1.00 for input and $4.05 for output, Small is roughly 70% cheaper on both counts. However, this service is still in beta, and the company does not yet recommend it for high-load production use.
Fine-tuning is also currently 50% off for a limited time. For the 64,000-token version, pricing per million tokens is $1.44 for output generation and $1.73 for training. The large version costs $4.68 and $5.61 respectively. The price difference is clear, but no end date has been given for the discount. When adopting the model, companies should also factor in the possibility of prices reverting to standard rates.
Usage Terms Aren't Fully Determined by Apache 2.0 Alone
The model card and Hugging Face page list Inkling-Small's license as Apache 2.0. However, the usage terms aren't fully covered by Apache 2.0 alone. Thinking Machines has established a separate "Model Acceptable Use Policy," which applies to the model's weights, parameters, related materials, and any modified versions. Obtaining or using the weights is treated as agreement to this policy.
This policy restricts things such as full automation of decisions that affect rights, unauthorized surveillance, and the creation of malicious code. Thinking Machines reserves the right to update the policy, and continued use after an update is treated as agreement to the new terms. Companies looking to integrate the model into their own systems need to separately verify both the technical feasibility of storing the weights and the legal question of whether their intended use complies with the policy.
The terminology is also more accurately described as "open-weight" rather than "open-source AI." The Open Source Initiative's Open Source AI Definition 1.0 requires not only the freedom to use, study, modify, and share for any purpose, but also detailed information about training data and complete training and execution code. Thinking Machines' documentation on training data is a general description shared across all of its services—stating that public information, third-party data, and synthetic data are used—but it does not provide an Inkling-Small-specific data inventory or complete training procedures.
The value of Inkling-Small lies not in bringing large-model capability down to consumer-grade hardware, but in bringing it one step closer to the range of computing infrastructure that enterprises can actually manage. While it surpasses the large version in coding and tool-use tasks, the gap widens on evaluations that directly probe knowledge. What should ultimately determine adoption decisions isn't the aggregate score. The next things worth verifying are whether this pattern of strengths and weaknesses holds up when tested against your own organization's specific tasks, and whether a 180GB+ execution environment or a beta-stage API will prove more stable for your needs.