Thinking Machines Lab

Overview

AI lab whose first production model, Inkling, shipped on 2026-07-15 as an open-weights release under Apache 2.0 (source).

The lab's stated positioning for that model is unusual and worth recording as the thing that distinguishes it: it explicitly does not claim Inkling is the strongest overall model, describing its role instead as a balanced, multimodal, customizable base for downstream fine-tuning (source). The product around it follows from that — same-day fine-tuning on the lab's own Tinker platform, and day-zero support in SGLang, vLLM, llama.cpp and Unsloth (source).

Key People

unknown — no leadership detail was obtainable from this environment; the lab's own site was not readable when this page was created.

Models & Products

• Inkling — released 2026-07-15. 975B total / 41B active multimodal MoE, 1M context, Apache 2.0 (source)
• Inkling Small — a smaller variant. It appears as its own row on the Artificial Analysis leaderboard read 2026-08-02 (Intelligence Index 40, Cost per Task USD $0.07) but no announcement for it was read, so it has no page (source)
• Tinker — the lab's fine-tuning platform; Inkling was fine-tunable on it from release day (source)

Recent Activity

• 2026-07-15: Inkling released — the lab's first production model and first open-weights release: 975B total parameters, 41B active per token, 1M-token context, native reasoning across text, images and audio, controllable thinking effort. Apache 2.0, full weights on Hugging Face in BF16 and NVFP4. Pretraining covered 45 trillion tokens of text, images, audio and video; the MoE design largely follows DeepSeek-V3 (256 routed experts, 2 shared, 6 active per token). → Inkling (source)

• 2026-08-02: Named as one of three open models on the Pareto frontier — Interconnects' "Latest open artifacts (#23)" names Inkling alongside Laguna S 2.1 and Kimi K3 (source). This is also how the release reached this wiki, 19 days after the fact.

Strategic Position

• Sits in the open-weights tier with a Western address, alongside Poolside and Mistral AI and against the Chinese open-weight labs (DeepSeek, Moonshot AI, Alibaba / Qwen AI Lab, Z.ai)
• The distinguishing bet is customizability over headline capability: a base sold for fine-tuning rather than for leaderboard position, with the fine-tuning platform as the product (source)
• Third-party measurement is consistent with the modest self-positioning. On the Artificial Analysis leaderboard read 2026-08-02, Inkling scores 41 on the Artificial Analysis Intelligence Index against Claude Opus 5 (max) at 61 and Kimi K3 (max) at 57, while its Median Tokens/s of 86 is well above both (source)

Related

• Inkling
• Open-Weights Policy Fight
• Poolside — the other Western open-weight release this window