Inkling and the fine-tuning bet: what an open model built for customization means for your business

On 15 July 2026, Thinking Machines Lab, the AI company founded by former OpenAI CTO Mira Murati, released its first model. Inkling is open-weight under an Apache 2.0 license, which permits commercial use, and it is deliberately not sold as a finished chatbot. The company's own materials say it is "not the strongest model available today, closed or open." It is sold as a starting point: a base model that organizations fine-tune on their own data and workflows. If you run a business and are weighing an AI API subscription against something built around your own operations, this release is worth ten minutes of your attention — not because you should use Inkling, but because of the decision it signals.

What happened

Inkling is a mixture-of-experts model with 975 billion total parameters, of which about 41 billion are active on any given token. It accepts text, images, and audio as input, produces text, and supports a context window of up to one million tokens. It was pretrained on 45 trillion tokens and released with full weights on Hugging Face, a model card, and support in the standard open-source inference stacks. A smaller sibling, Inkling-Small, was previewed but its weights have not shipped yet.

Two details in the model card matter more to a business reader than any benchmark. First, the license: Apache 2.0, the same permissive license used by much of Mistral's and Alibaba's Qwen lineup, meaning a company can download, modify, fine-tune, and deploy it commercially without negotiating terms. Second, the hardware: the full-precision checkpoint requires a GPU cluster with at least 2 TB of combined VRAM, roughly sixteen NVIDIA H200s, and even the compressed NVFP4 version needs at least 600 GB across four to eight data-center GPUs. Thinking Machines expects most users to reach the model through its Tinker fine-tuning platform or through hosted providers such as Together AI, Fireworks, Databricks, and Baseten, not to run it themselves.

The release lands in the middle of an argument the industry is having out loud. Two days earlier, Microsoft CEO Satya Nadella warned that enterprises using proprietary AI models effectively pay twice: once in subscription fees, and again by handing over the business knowledge embedded in their prompts and corrections, which can be absorbed into future model versions. And in late June, Thinking Machines published joint results with Bridgewater Associates in which an existing open model, trained further on the fund's own financial expertise, scored 84.7% on financial-reasoning tests — ahead of top proprietary models, at roughly one-fourteenth the running cost. That result comes from the two companies' own evaluation, not an independent one, so treat it as a signal rather than a proof. But the direction of the signal is clear: on a narrow, well-defined domain, a mid-sized model tuned on your data can beat a frontier model that has never seen your business.

Why it matters for operators

The first practical lesson is uncomfortable for anyone who equates "open weights" with "run it in the back office." A 975-billion-parameter model with a 600 GB minimum VRAM footprint is not something a restaurant group, a clinic chain, or a retail operator will ever host on-premise. The open-model market has split: at one end, genuinely self-hostable models in the 7B–70B range that run on a single serious GPU; at the other, frontier-scale open models like Inkling, DeepSeek, and Kimi that are "open" in license but data-center-class in practice. For most Saudi businesses, the realistic ways to touch a model like this are a hosted API or a managed fine-tuning platform — which means the data-control questions do not disappear just because the license is permissive.

That leads directly to the second lesson. Fine-tuning a model on your operations means shipping your operational data — orders, tickets, contracts, HR records — to wherever the fine-tuning happens. For a Saudi business, that is a PDPL question before it is a technology question. Personal data in a training set sent to a US-hosted platform needs the same cross-border transfer review as any other processing arrangement, and some data, such as employee records or customer identities, may be better pseudonymized or excluded entirely. The build-vs-buy decision for custom AI now has three lanes, not two: rent a general API, fine-tune on someone else's platform, or fine-tune and host within infrastructure you control. Each lane has a different cost, capability, and compliance profile.

Third, for Arabic-first operations the honest answer is: unverified. The model card claims "English, general multilingual support" and reports a respectable 88.7% on Global-MMLU-Lite, but publishes no Arabic-specific evaluation, no dialect coverage, and no Arabic examples. That is not a criticism — it is simply a gap you should test before committing anything customer-facing to it. Regionally trained options like ALLaM, which we covered in June, still hold the stronger verified claim on Arabic.

Where this helps, and where it is premature

The genuinely useful idea in this release is not the model. It is the product category it is trying to establish: fine-tuning as a routine business activity rather than a research project. The Bridgewater pattern — take an open model, train it on a narrow domain where your firm has real expertise, run it at a fraction of frontier-API cost — maps directly onto operational software. The candidates are the boring, high-volume tasks where the rules live in your staff's heads: coding supplier invoices to the right accounts, triaging support tickets by urgency and department, drafting the weekly operations summary in your house format, extracting order details from voice notes. These are exactly the tasks where a general chatbot is mediocre, because it has never seen your categories, your exceptions, or your Arabic-English code-switching.

It is premature, though, for most businesses to fine-tune anything this quarter. Fine-tuning still requires machine-learning judgment: building an evaluation set, generating or labeling training data, and knowing when the tuned model is actually better rather than differently wrong. TechCrunch's assessment — that this "requires serious machine learning talent" — matches our experience. It is also premature to bet on any single model. Inkling shipped yesterday; its smaller sibling has no weights yet; Qwen's next open generation is still pending; prices and licenses shift monthly. A business that hard-codes its stack to this week's model will be renegotiating with itself by October.

Cicada Solutions view

Our advice is to prepare for fine-tuned AI without buying it yet, because the preparation is valuable on its own. Start by picking one task, not a strategy: a single high-volume, low-ambiguity job where you can define what a correct answer looks like — an invoice coded to the right account either matches your accountant's decision or it does not. Then get the data ready: most businesses discover their real blocker is not model choice but the fact that the examples a model would learn from are scattered across WhatsApp, spreadsheets, and a POS export. A system that captures those decisions cleanly — the ticket, the category a human chose, the correction when they got it wrong — is useful today for reporting and becomes training data tomorrow. Build the evaluation set now too: fifty to a hundred real examples with known correct answers, which is the same artifact you would need to judge any AI vendor's demo. Run the PDPL review before any data leaves your environment, and prefer designs where personal identifiers never enter the training set at all. And keep the model swappable: whatever you build should treat the model as a replaceable part behind an interface, because the one certainty in July 2026 is that the best option in January 2027 will be different. What changed this week is not that your business suddenly needs a 975-billion-parameter model. It is that the industry's serious money is now betting that the durable value in AI sits in your data and your workflows, not in anyone's chatbot subscription. That is a bet operators can quietly get ready to win.

Sources

• Thinking Machines Lab: Inkling — Our open-weights model - published 2026-07-15, accessed 2026-07-16.
• Inkling Model Card, Thinking Machines Lab - published 2026-07-15, accessed 2026-07-16.
• Hugging Face: thinkingmachines/inkling - accessed 2026-07-16.
• Thinking Machines Lab: Learning to Replicate Expert Judgment in Financial Tasks - published 2026-06, accessed 2026-07-16.
• TechCrunch: Thinking Machines amps up its bet against one-size-fits-all AI with its first open model, Inkling - published 2026-07-15, accessed 2026-07-16.