On 15 July 2026, Thinking Machines Lab announced Inkling and Inkling-Small preview. The lab’s first model combined native text, image and audio reasoning with controllable thinking effort and a fine-tuning platform.
This release brief was checked against first-party material on 10 August 2026. The date above is the public announcement date, not the date a repository was created or a third-party provider added the model. Where access or weights arrived later, that distinction is recorded below.
Release record
| Field | Verified detail |
|---|---|
| Announcement | 15 July 2026 |
| Availability or weight release | 15 July 2026 as full weights, Tinker fine-tuning access and an Inkling-Small preview |
| Release type | open-weight general-purpose multimodal foundation-model family |
| Access | full Inkling weights, Hugging Face recipes and hosted Tinker customisation; licence terms must be checked directly |
| Architecture | a 975B-total, 41B-active MoE transformer with 256 routed experts, two shared experts and interleaved sliding/global attention |
| Maximum stated context | one million tokens in the model, with 64K and 256K options initially exposed through Tinker |
What changed
Inkling is one of the largest new open-weight foundations in the window. It was pretrained on 45 trillion tokens spanning text, image, audio and video, and it exposes effort as an explicit cost-performance control. Thinking Machines also released a smaller preview and recipes for customisation. This makes it valuable as a base-model release even though the lab plainly says it is not the strongest model overall.
The practical comparison is therefore not simply whether Inkling and Inkling-Small preview has the largest headline score. Teams need to ask whether its architecture, access terms, latency, tool behaviour and evaluation setup match the workload they actually intend to run. A model can lead one harness while losing on cost, refusal behaviour, multilingual quality or repeatability in another.
Benchmarks worth retaining
| Evaluation | Reported result | How to read it |
|---|---|---|
| Design Arena Agentic Web Dev | 1,257 Elo in the launch table | Blinded artifact preference, close to several frontier models |
| Terminal-Bench efficiency | matched Nemotron 3 Ultra at roughly one-third the tokens | Effort-curve claim rather than a single fixed score |
| Pretraining scale | 45T multimodal tokens | Training disclosure, not a capability benchmark |
These are release-time results, not independently reproduced guarantees. Several multimodal results were produced on a different pre-release checkpoint, and the 1M architectural limit is not the same as the shorter contexts initially available through Tinker. Scores should remain attached to the disclosed effort setting, agent harness, tool access, timeout, context-management policy and judge model. Moving a number into a procurement sheet without those conditions creates false comparability.
Architecture and access
Inkling and Inkling-Small preview is described as a 975B-total, 41B-active MoE transformer with 256 routed experts, two shared experts and interleaved sliding/global attention with one million tokens in the model, with 64K and 256K options initially exposed through Tinker of stated context. Its access position at verification time is full Inkling weights, Hugging Face recipes and hosted Tinker customisation; licence terms must be checked directly. That wording matters: open weights, source-available weights, an API, a product preview and a research demonstration give adopters very different rights and different levels of reproducibility.
Before deployment, record the exact model identifier or checkpoint, inference stack, quantisation, reasoning setting, region, price schedule and supplier terms. If the release uses a custom licence, read the licence itself rather than relying on the word “open” in launch copy. If it is API-only, preserve the dated documentation and change-notice route because the served snapshot can change without a downloadable artefact.
What an evaluation should test next
For Inkling and Inkling-Small preview, a credible internal gate should include:
- a frozen set of representative tasks with pass, fail and abstain criteria;
- a matched baseline using the same tools, timeout, prompt budget and reviewer rubric;
- repeated runs to expose variance rather than reporting a single best attempt;
- latency, token use and total task cost alongside task success;
- adversarial, multilingual and long-context cases relevant to the real deployment; and
- rollback evidence showing the previous model can be restored safely.
The wider model change-control guide explains how to keep model, prompt, tool and corpus changes reconstructable. The AI dependency inventory guide covers the release and supplier records needed after deployment.
AIEngine verdict
Inkling is a major open-model debut with unusually clear architectural disclosures. Teams should benchmark the downloadable checkpoint separately from Tinker and preserve the exact effort setting because token use is part of its performance claim.
This is a launch assessment, not a certification. Benchmark leadership is useful evidence of where to test; it is not authorization to place the model in a high-impact workflow without domain evaluation, security review and an accountable owner.
Primary sources
Image provenance
Hero image: Thinking Machines Lab official release artwork. The locally served WebP is a crop of the first-party release or model-card asset recorded in the repository provenance manifest.
