Thinking Machines Lab launches Inkling: 975 billion open-weight parameters

When Mira Murati left OpenAI in September 2024 — after years as CTO and architect of much of the company's product strategy — the obvious question was: what is she going to build? The answer arrived with more weight than many expected. Thinking Machines Lab has just launched Inkling, an open-weight language model based on a mixture-of-experts (MoE) architecture with 975 billion total parameters and 41 billion active parameters per inference. This is not a lab experiment. It is a positioning statement in a market that, until recently, seemed reserved for OpenAI, Google, and Anthropic.

What Inkling is and how it works

Inkling adopts the MoE architecture — the same logic underlying GPT-4 and Gemini 1.5 — but at a scale that places the model in the frontier tier. In practice, MoE means the model does not use all 975 billion of its parameters at once. With each generated token, only a subset of "experts" is activated — in Inkling's case, around 41 billion parameters. This has a direct and highly relevant consequence for those operating infrastructure: inference costs drop dramatically compared to a dense model of equivalent size.

The fact that it is open-weight is the other structural differentiator. Unlike a closed API, an open-weight model can be downloaded, fine-tuned, and hosted on proprietary infrastructure. Companies with compliance restrictions, sensitive customer data, or simply a desire for control over latency and cost per token now have a cutting-edge alternative — without depending on a single vendor.

Why size matters here

975 billion total parameters puts Inkling in the same conversation as Mistral's Mixtral 8x22B and DeepSeek-V3, both reference MoE models in the open segment. The relevant difference is origin: Thinking Machines Lab arrives with high-profile capital, a team composed of ex-OpenAI members, and the weight of Murati's track record as one of the most influential figures in building ChatGPT. This matters because training quality — the data, the RLHF process, the alignment — is just as decisive as the parameter count.

What changes in the model market

Until 2023, the conversation about frontier models was binary: you either used GPT-4 via API or settled for clearly inferior alternatives. That landscape changed with Meta's Llama 3, China's DeepSeek, and France's Mistral. Inkling deepens that disruption.

For companies using AI in product, support, or engineering, the impact is concrete:

• More leverage in price negotiations. When three or four frontier models are competing for the same workload, vendors like OpenAI and Anthropic feel pressure to adjust pricing and terms. We have already seen this happen — GPT-4o mini and Claude Haiku exist, in part, because of this competition.
• More licensing options. Open-weight is not the same as pure open-source, but it does allow on-premise or private VPC deployment. For a bank, a healthtech company, or any business with regulated data, this can eliminate a real blocker.
• Benchmarks are no longer the only criterion. With more models of comparable quality, the decision shifts to latency, cost per token, fine-tuning capability, and support. This favors companies that build internal model evaluation expertise — not just those who subscribe to the most expensive API.

A use case to anchor the decision

Imagine a mid-sized e-commerce company that uses LLMs for three functions: product description generation, support ticket triage, and response suggestions for human agents. Today, that stack likely runs on GPT-4o or Claude 3.5, with monthly costs that can range from R$ 8,000 to R$ 40,000 depending on volume.

With a model like Inkling running on proprietary infrastructure — via dedicated cloud or even on-premise hardware for larger companies — inference costs drop to the price of compute, with no vendor margin. Fine-tuning allows the model to learn the company's tone and catalog without relying on massive prompts. And latency can be optimized according to the internal architecture.

This is not fiction: it is exactly what mid-sized European and North American companies are already doing with Llama 3 and Mistral. Inkling enters that conversation with higher quality parameters.

What we still don't know — and need to know

Any honest assessment requires caveats. Inkling has been announced, but independent benchmarks and head-to-head comparisons on real-world tasks are still being conducted by the community. The quality of reasoning in Portuguese — relevant for any Brazilian company — depends on the training data mix, which Thinking Machines Lab has not yet detailed publicly.

Furthermore, running a model with 975 billion total parameters requires serious infrastructure. Even with MoE reducing inference costs, the memory footprint for loading the weights is substantial. For most SMBs, the practical route will be via API — if and when Thinking Machines Lab makes a commercial endpoint available — or through cloud providers that host the model.

The landscape that is taking shape

The launch of Inkling confirms a trend I have been closely following: the AI frontier is no longer the exclusive property of two or three American labs. This is structurally good for any company that uses AI as a strategic input — because competition produces better models, lower prices, and more options for how and where to run them.

For technology and product managers in Brazil, the practical message is simple: the time to build internal competency for evaluating, comparing, and switching models as the market evolves is now. Those who became locked into a single vendor in 2023 have already felt the cost of that. With Inkling on the table, the options have gotten better — and the excuse for not diversifying, smaller.