Thinking Machines Lab released Inkling — a 975B open-weights MoE built for customization, not crown jewels. Continuity still beats the hype cycle.
Via Thinking Machines Lab: Inkling: Our open-weights model
Thinking Machines Lab released Inkling on July 15, 2026 — a multimodal Mixture-of-Experts model with 975 billion total parameters and about 41 billion active, a context window up to 1 million tokens, and full weights on Hugging Face. It was pretrained on 45 trillion tokens of text, images, audio, and video.
Here is the part that matters for an owner-led firm: the company says Inkling is not the strongest overall model available, open or closed. They built it as a customization base — multimodal, controllable thinking effort, and fine-tunable on their Tinker platform. That is a Model Selection story, not a coronation.
If your staff already bounce between chat tools and nobody owns continuity, another impressive open-weights release does not fix the gap. It widens the menu. Someone still has to decide which model is allowed for which workflow — and what happens when the next release lands.
Thinking Machines Lab — founded by former OpenAI CTO Mira Murati — is betting that organizations will adapt models for themselves rather than rent a single frontier assistant forever. Whether that bet pays their Tinker business is their problem. Whether your firm adopts the newest base without a fallback is yours.
Inkling reasons natively over text, images, and audio, with a dial for thinking effort so developers can trade tokens for quality. Thinking Machines reports it can match Nemotron 3 Ultra on Terminal Bench 2.1 at roughly one third of the tokens at comparable score — useful if you care about cost and latency, not just max-effort leaderboard screenshots.
They also previewed Inkling-Small (276B total / 12B active), which early numbers show close to the larger model on many reasoning and agentic tasks. Full Small weights come after testing finishes. For firms that run models millions of times inside longer workflows, a controllable effort curve is more honest than a single headline score.
Open weights still have a catch. Downstream coverage notes the BF16 checkpoint needs on the order of 2 TB of aggregate GPU memory; an NVFP4 checkpoint drops that toward roughly 600 GB. "Downloadable" is not the same as "cheap to run in-house." Inference partners already listed include Together AI, Fireworks, Modal, Databricks, and Baseten — plus Hugging Face transformers support. Jurisdiction, residency, and vendor continuity still sit on the SME owner's desk.
Thinking Machines frames the bet clearly: customization beats one-size-fits-all. Fine. Your policy still needs a named fallback when the preferred base model changes price, availability, or safety behavior after fine-tuning.
They also showcase agentic coding, Design Arena web-app rankings among strong open-weights peers, and safety numbers on FORTRESS that they claim lead the open-weights set they compared. Useful context — still not a reason to declare a firmwide default without a job map.
Treat Inkling like any other model-landscape event — evaluate it against jobs, not vibes.
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.
This release maps to Model Selection & Continuity Planning. The decision is not "is Inkling good?" It is "which model fits which job, at what cost and residency, with what fallback when the landscape moves again?"
AgentsROI stays vendor-neutral. I help owner-led firms in law, accounting, health, and wealth choose deliberately — cloud, hybrid, or local — and keep a continuity plan so a new open-weights drop does not become an accidental production dependency. If your team is already improvising with the newest model of the week, a Shadow-AI Risk Assessment finds the gap between policy and practice.
Primary next step: get a model map that assumes releases like Inkling will keep arriving — because they will.
Thinking Machines opened the weights and invited customization. That is useful for builders. For an SME owner, the useful move is still the boring one: assign models to jobs, cap cost and data exposure, and keep a fallback.
Ready to pressure-test your model map? Start with a Model Selection & Continuity review at AgentsROI.ai.
This article summarizes publicly reported information and is for general informational purposes only. It does not constitute legal, tax, financial, investment, security, or compliance advice. AgentsROI.ai is not a law firm, accounting firm, or registered investment adviser. Facts, pricing, statistics, and product capabilities cited here reflect the sources listed at the time of writing and may change. Readers should verify current information independently and consult qualified professionals regarding obligations specific to their industry, jurisdiction, and circumstances—including applicable New York State and New York City requirements. AgentsROI.ai may have commercial relationships with vendors mentioned; where material, such relationships are disclosed. Nothing in this article is an endorsement of any specific AI product, model, or provider.
