The Lab Notes

Thinking Machines

Thinking Machines releases Inkling 2 Small weights under Apache 2.0

Thinking Machines made the Inkling 2 Small weights downloadable under Apache 2.0, and the main checkpoint needs at least 500GB of GPU memory.

A glass inkwell half full of dark ink with a green fountain pen nib rising from its neck, outlined in pale yellow, set between a black background with a deep blue glow and thin gray arcs on the left and a yellow background tiled with rows of pink and blue inkwells with green nibs on the right.

Thinking Machines Lab has released the weights of Inkling 2 Small for download under the Apache 2.0 license, so companies can run and fine-tune the model on their own hardware. The weights are generally available now, not in preview.

The company introduced the model Nov. 5 with closed weights, available only through Tinker, its fine-tuning platform. It said the weights would follow “pending further safety evaluations by us and our partners.”

The Weights Are on Hugging Face Under Apache 2.0

Thinking Machines posted the weights on Hugging Face, and its Nov. 12 update to the launch post says they are “now available for download on HuggingFace.” The model card lists the license as Apache 2.0, the one the company named Nov. 5. The company said it ran “a wider suite of evaluations, with more external testers, than any previous release” before the release.

The Main Checkpoint Needs 500GB of GPU Memory

The main checkpoint needs at least 500GB of combined GPU memory, such as 2 Nvidia B300 or 4 Nvidia H200 GPUs. It is stored in NVFP4, the precision the model was trained in, and the company recommends it for all uses. A second checkpoint in MXFP8 needs at least 800GB, such as 4 B300 or 8 H200, and gains “a very marginal 1–2 percentage points.”

The model is a mixture-of-experts design with 594 billion total parameters and 17 billion active per token. Each token is routed to 8 of 384 experts across 48 layers. It takes text and images, returns text, and handles up to 1 million tokens of context. It does not support audio.

Local inference runs on SGLang, vLLM, TokenSpeed and Unsloth. Post-training runs on Miles, SkyRL and Snowflake Cortex Training.

Hosted Access Costs $0.30 per Million Input Tokens

OpenRouter lists Inkling 2 Small at $0.30 per million input tokens and $1 per million output tokens, plus a free variant, as of Nov. 13. Nvidia, Fireworks, Together, Baseten, Nebius, Modal and Cloudflare also host it through their own APIs.

The Benchmark Scores Are Thinking Machines’ Own

In its Nov. 5 post, Thinking Machines reported a score of 30.6 on the Artificial Analysis Intelligence Index v5.0. It also reported 86.3% on SWE-Bench Pro v2 (Hard), 81.7% on ARC-AGI-2 and 90.2% on BrowseComp with compaction.

The model card’s comparison table puts the Intelligence Index v5.0 scores at 30.6 for Inkling 2 Small, 29.1 for MiniMax-M3.1-Flash, 27.8 for GLM-5.3-Flash and 26.7 for MiMo V2.6 Flash. All of these figures come from the company.

The company said Inkling 2 Small “is part of a ladder of models; we look forward to sharing more rungs with you soon.”

Our View

We think the license and the memory floor have to be read together. Apache 2.0 lets any company download the weights, fine-tune them and run them on its own hardware. The 500GB floor sets what that hardware is: a server with 2 B300 or 4 H200 GPUs.

A company that already operates machines like that can fine-tune and serve the model itself. A developer without them cannot load the main checkpoint and will reach the model through a host such as OpenRouter at the dated price.

Owning the weights pays off for a team that already runs, or will rent, a server with at least 500GB of GPU memory, and only if the scores Thinking Machines reported hold up on that team’s own work.