Inkling 2 Small: Hosted on Tinker Now, Open Weights Planned Next Week

As of its November 5 announcement, Inkling 2 Small is initially available through Thinking Machines Lab's own Tinker platform the November 5 announcement. This is a hosted API service; the model's weights are currently closed and cannot be downloaded. You can try the model in the Tinker Playground and use it for inference and fine‑tuning via the API. The Cookbook documentation provides the tml‑renderers package and guidance for integrating Inkling models into workflows the Cookbook guide.

In its November 5, 2026 announcement, Thinking Machines Lab states its intent to release the full model weights under an Apache 2.0 license the November 5 announcement. This release is pending the "completion of external safety testing and red‑teaming" and is intended for "next week" relative to that announcement. The release of Inkling 2 Small’s weights is also pending further safety evaluations by the company and its partners. Once released, developers will be able to download, self‑host, and modify the model. Checkpoints for fine‑tuning on Tinker will also become downloadable at that time.

The model's architecture is a Mixture‑of‑Experts transformer with 594 billion total parameters and 17 billion active per token the model card.

Context Variants: Two context lengths are listed on Tinker:

• 64K tokens (thinkingmachines/Inkling‑2‑Small)
• 256K tokens (thinkingmachines/Inkling‑2‑Small:peft:262144) Tinker pricing page.

The model card states a maximum context window of up to 1 million tokens the model card, but the currently served variants are 64K and 256K.

Input/Output Modalities:

• Accepts: Text and images (via a 292M‑parameter vision encoder).
• Does NOT accept: Audio. The model card for Inkling 2 Small does not list audio as a supported modality, unlike its predecessor Inkling‑Small the model card.
• Outputs: Text only.

Tinker charges different rates for prefill (processing input), sampling (generating output), and training (forward/backward pass). Prices are per million tokens and are lower for cached prefill tokens Tinker pricing page.

For the standard 64K context model (thinkingmachines/Inkling‑2‑Small):

• Prefill: $0.30 (Cached: $0.03)
• Sample: $1.00
• Train: $1.50

For the 256K context variant (thinkingmachines/Inkling‑2‑Small:peft:262144):

• Prefill and sample prices are identical to the 64K version.
• Train: $3.00 (The cost increases with the longer context.)

Hypothetical Cost Example: An inference workload processing 10 million prefill tokens (with a 50% cache hit rate) and generating 2 million new tokens would incur approximate costs of: (5M * $0.30) + (5M * $0.03) + (2M * $1.00) = $1.50 + $0.15 + $2.00 = $3.65 for the prefill, cached prefill, and sampling components. This illustrates the aggregate inference cost across multiple requests, not a single unsupported context request or a training/fine‑tuning cost.

Your choice depends on timeline, control needs, and budget.

Use Tinker Hosted Access Now if:

• You need to evaluate or build with Inkling 2 Small immediately.
• You prefer a managed API without infrastructure overhead.
• Your usage aligns with the per‑token pricing and you can work within the 64K or 256K context limits.

Wait for Open Weights if:

• You require full control, need to modify the model, or run it on your own hardware.
• Your application demands the full 1M token context (once implemented).
• Your cost structure favors self‑hosting over per‑token API fees.

For comparing Inkling 2 Small's capabilities and pricing against other models, the OpenTools model catalogue is a useful resource.