Inkling: a 975B open MoE built to be fine-tuned, not to top leaderboards
Thinking Machines Lab (founded by ex-OpenAI CTO Mira Murati) released Inkling on 2026-07-15: a 975B-parameter mixture-of-experts model (41B active per token, up to 1M-token context, trained on 45T multimodal tokens of text/image/audio/video), open-weights under Apache 2.0, with same-day fine-tuning on its Tinker platform. Thinking Machines itself says Inkling "is not the strongest overall model available today, open or closed," positioning it instead as an open Western base for customization; its MoE design "largely follows DeepSeek-V3" and it is competitive with, but not consistently superior to, Chinese open models (Kimi K2.5/K2.6, GLM 5.2). Note: the Apache 2.0 license and full spec live on the model card, not the cited news post.
Why it matters (spine). The competitive frame is shifting from "who tops the leaderboard" to "who owns the best base to specialize." Thinking Machines is explicit that Inkling is not the strongest model available; it is pitched as an open, Apache-2.0 Western foundation you fine-tune on your own data, with same-day adaptation on its Tinker platform. When the base model is a commodity and specialization is a same-day service, the durable, defensible input becomes the proprietary, verified ground truth you feed it — which is exactly what a knowledge garden is.
What would prove it wrong (falsifier). Independent tests show Inkling is no more adaptable or cheaper to specialize than existing open bases, collapsing the "customization over leaderboard" thesis.
Live-status check + triangulation. Quoted from the Inkling model card (Thinking Machines Lab, dated 15 July 2026): "Inkling is not the strongest overall model available today, open or closed. Instead, a combination of qualities makes it a good open-weights base for customization." Fetched from the primary model card in-session, and corroborated by Simon Willison (simonwillison.net, 16 Jul 2026), VentureBeat, and Latent Space, plus the model card itself.
Correction carried forward (discipline). The numeric and naming claims check out against the primary — total 975B, active 41B per token, 1M-token context, and 45T multimodal training tokens all match (unlike the sibling GPT-5.6/5.4 card, this one is clean). Two provenance caveats: the Apache 2.0 license is supported but appears on the model card, not the originally-cited news URL, so that URL alone is insufficient to verify it; and the "Mira Murati" attribution is correct (she founded Thinking Machines Lab) but her name does not appear on the linked release page — it comes from independent coverage.
Treat Inkling as a candidate private substrate for a garden. Fine-tuning-as-a-service makes verified proprietary ground truth the scarce input, and that input is RDKG.
domains: machine-layer · open-weights · mixture-of-experts · fine-tuning · multimodal · moat