Thinking Machines opens Inkling 2 Small weights for deployment Thinking Machines opens Inkling 2 Small weights for deployment

A download changes the deployment choice

The November 12 update links the public download and several serving integrations. An existing evaluation can now compare an installation the team operates with the hosted route it could already use. This adds an operating choice without requiring customers to leave Tinker.

Small is a model name

The model card gives two hardware paths. It lists at least 500GB of combined GPU memory for the NVFP4 checkpoint and 800GB for MXFP8. Example configurations include four H200 GPUs for the first and eight for the second. Those are deployment requirements published by the lab, not measurements from our own hardware. The active parameter count describes the part used during processing; it does not replace the memory requirements for the complete model. For a procurement decision, the useful starting point is the actual checkpoint and serving configuration, not the word Small.

The serving recipe narrows the promise

The vLLM recipe, updated November 12, separates configurations by context length. Its verified GB200 setup supports 262,144 tokens on one four-GPU tray and the full 1,048,576-token context across two trays. The recipe also specifies a dedicated container image while release support is pending. This is a concrete difference between a model capability and an installation setting. A service configured for the shorter limit cannot accept the longer request merely because the model card advertises it. The deployment contract includes both the checkpoint and the runtime that loads it.

Throughput comes with a workload

Inferact and Thinking Machines published performance results with the vLLM integration on November 12. Their agent workload reused long contexts through prefix caching, appended tool output and generated 1,024 tokens per turn. Reported throughput changed as session context grew, even with a fixed target of 100 output tokens per second per user. These details make the benchmark useful as a configuration example. They also explain why a single headline speed cannot price an application: request length, repeated material and simultaneous sessions determine which part of that example resembles its traffic.

Ownership includes the operating work

The lab’s July 31 release policy describes staged access as a way to prepare the surrounding ecosystem before weights become public. That earlier policy explains the sequence; November 12 is the date this checkpoint became downloadable. The practical choice now splits into separate questions: whether the application benefits from controlling the weights, whether its infrastructure can serve them, and whether its tests support the intended use. An open checkpoint answers the access question. It leaves the operator to establish a release process for the application built around it.