Thinking Machines' open Inkling models: what the 276B Inkling-Small offers

Thinking Machines Lab’s Inkling family now spans two released open-weight models: the 975-billion-parameter flagship Inkling and a lighter Inkling-Small with 276 billion total parameters and 12 billion active per token. Both are licensed under Apache 2.0, reason over text, images and audio, and can be fine-tuned on the company’s Tinker platform. Inkling-Small’s specifications in this article are those of the July 2026 release; specifications for any newer Inkling 2 Small model could not be confirmed against the source material for this report and are not given here.

For builders, the useful question is what the small model gives up against the flagship and what it gains. The record so far points to a model about a quarter of the size that lands close to Inkling on several third-party measures, with some disagreement between listings on context length and input types.

How Thinking Machines launched the Inkling family

Latent Space’s AI News roundup dated July 16, 2026 describes Inkling as Thinking Machines Lab’s first fully released open-weights foundation model, positioned as a customisable multimodal base model rather than a benchmark-maxed flagship. It quotes the company’s announcement:

Our model, called Inkling, is a Mixture-of-Experts transformer with 975B total parameters, 41B active. It supports a context window of up to 1M tokens. It was pretrained on 45 trillion tokens of text, images, audio and video.

A mixture-of-experts (MoE) model splits its feed-forward layers into many “expert” sub-networks and routes each token through only a few of them, so the active parameter count, not the total, drives the compute per token. The same announcement previewed Inkling-Small as “a lighter-weight model with 12B active parameters, trained with a similar recipe, that achieves strong performance with even lower cost and latency”.

According to the roundup, Mira Murati called Inkling the company’s “first model,” “trained from scratch,” with same-day fine-tuning on Tinker, and John Schulman said pretraining began the previous winter, with a small team building coding, reasoning and agentic training on top from mid-January. Staff stressed that this was a day-one release and a foundation for later iterations, not the company’s final frontier push.

Sources disagree on the flagship’s release date. Latent Space reports the launch on July 16, while LLM Stats lists Inkling as released on 2026-07-21. Both LLM Stats and the Sulat model listing give Inkling-Small a release date of July 30, 2026.

What Inkling-Small offers next to the flagship

The Sulat listing describes Inkling-Small as an MoE transformer with a 42-layer decoder and a hidden size of 4096. It uses hybrid attention: a 512-token local sliding window, with every sixth layer attending globally, and the same sparse feed-forward routing as the larger checkpoint. That matches the flagship design that community readers extracted, a 5:1 local-to-global layer ratio with window size 512.

Spec	Inkling	Inkling-Small
Total parameters	975B	276B
Active parameters per token	41B	12B
Inputs (maker’s description)	Text, images, audio	Text, images, audio
Output	Text	Text
Licence	Apache 2.0	Apache 2.0
Released	July 16 or 21, 2026 (sources differ)	July 30, 2026

Other details in coverage summarised on the Sulat page: Inkling-Small was trained on NVIDIA GB300 NVL72 systems, and both original and NVFP4-quantised checkpoints are offered for deployment on Blackwell hardware. NVFP4 is a 4-bit number format that cuts memory needs, which coverage says lowers the barrier to running the model locally. The model also supports variable thinking effort, from “minimal” to “xhigh,” tied to compute consumption so users can trade cost against quality per task.

Vision and tool use for agentic work

The Inkling line’s multimodal backbone is the main thing that sets it apart from text-only peers, according to the Sulat listing. Thinking Machines highlights Inkling-Small for cropping, zooming and programmatic image inspection, according to coverage summarised on the same page. In practice, this means an agent can examine parts of an image through tool calls rather than taking it in as a single fixed view.

The listing marks Inkling-Small as supporting reasoning, tool calling, structured output, attachments and temperature controls. In Latent Space’s account of the flagship launch, Tinker customers highlighted concise reasoning and strong tool calling over raw benchmark chasing.

The sources do not agree on which inputs the small model takes. The Sulat description says text, images and audio, but the Baseten provider entry on the same page lists only text and image input. LLM Stats lists text, images, audio and video for both models. Builders who depend on audio or video should test the specific endpoint they plan to use.

What the benchmarks say, and who says it

In Thinking Machines’ own reporting, Inkling-Small scored above 31% on Humanity’s Last Exam, against Inkling’s 29.7%, and crossed 80% on SWE-Bench Verified, according to coverage summarised on the Sulat page. The company says it is competitive with or better than peers in its weight class on Terminal-Bench 2.1 (agentic tool use), text-only HLE (reasoning) and IFBench (instruction following).

Third-party numbers point the same way, but with narrower margins:

• Artificial Analysis gave Inkling-Small a score of 40 on its Intelligence Index, within one point of the flagship’s 41, according to VentureBeat as summarised on Sulat.
• LLM Stats rates Inkling-Small at 37.7 on its composite score, against 36.6 for Inkling. On its 13 shared benchmarks, though, Inkling wins 7 and Inkling-Small wins 6.
• Per LLM Stats, Inkling-Small leads on GDPval-AA, IFBench, MCP Atlas, MMMU-Pro, SWE-Bench Verified and Terminal-Bench 2.1, while Inkling leads on AIME 2026, CharXiv-R, Global-MMLU-Lite, MMAU, SimpleQA Verified, Tau3 Banking and VoiceBench Avg.

The split makes sense. The smaller model does better on coding, agentic and tool-use tests, while the flagship keeps the edge on maths competition problems, factual recall, audio and some chart reading. LLM Stats also scores Inkling-Small higher on its agents index (22.2 vs 19.4) and coding index (28.6 vs 26.3).

For wider context, Latent Space reports that independent commentators called Inkling the strongest US-based open-weight release so far, while still placing it behind the top Chinese open-weight models and the best closed models on some benchmarks.

Access: Tinker, open weights and hosting partners

Full weights for Inkling-Small are on Hugging Face, alongside a Tinker Playground, and the model supports multimodal fine-tuning across text, image and audio, according to coverage summarised on Sulat. The open-weights release lets organisations self-host and fine-tune. The company’s launch post said Inkling-Small was available immediately for fine-tuning on Tinker.

The flagship launched with broad day-zero support across vLLM, SGLang, Modal, Baseten, Databricks and Hugging Face, according to Latent Space. LLM Stats lists Inkling-Small as served by Thinking Machines Lab and DeepInfra, and Sulat lists a Baseten endpoint under the key thinkingmachines/inkling-small. LLM Stats calculates Inkling-Small at about 3.3 times cheaper per token than Inkling on a blended basis.

Context length is the clearest conflict in the record. For the flagship, Latent Space reports that open-weights checkpoints support up to 1M tokens, while the Tinker and API context is described as 256K. For Inkling-Small, the Baseten listing on Sulat shows a context limit of 1,048,576 tokens and output of 32,768, while LLM Stats lists 256,000 tokens. The usable window appears to depend on where the model is served.

What it means for builders

The pitch for Inkling-Small is Inkling-class reasoning at roughly a quarter of the size. With 12B active parameters, each token costs far less compute than the flagship’s 41B, and VentureBeat frames the model as a fit for enterprises with moderate GPU resources, per the Sulat summary. Apache 2.0 is a permissive licence: it allows commercial use and modification without the regional or usage carve-outs some recent open-weight releases carry.

Some limits to keep in mind:

• Total size still matters for self-hosting. All 276B parameters must fit in memory even though only 12B run per token. The NVFP4 checkpoints help, but are aimed at Blackwell-class hardware.
• Listings conflict. Context length and supported input types differ between sources, so check the provider’s own limits.
• No knowledge cutoff is published. LLM Stats notes that neither model specifies one.
• The company frames its benchmark claims as competitive, not leading. Thinking Machines itself presented Inkling as a solid base for customisation rather than a frontier model.

What to watch next

Thinking Machines called Inkling “the first in a family of models of different sizes,” and staff described the release as a foundation for future iterations. Things to watch: whether any newer small Inkling model publishes a model card with its own parameter counts, context limits and licence terms; whether independent leaderboards such as Artificial Analysis and LLM Stats rank it against the July Inkling-Small; and whether its open weights ship on the same day as Tinker access, as Inkling’s did, or later.