Inkling: The 975B Open-Weight Model That Isn't Trying to Be the Best

Mira Murati left OpenAI in September 2024. Eighteen months later, her company just shipped its first model.

Inkling is a 975-billion-parameter open-weight model from Thinking Machines Lab, the San Francisco startup Murati co-founded with Lilian Weng. It’s backed by a16z. The company is valued at $12 billion. And the model is available under Apache 2.0 — meaning anyone can download it, modify it, and deploy it.

Here’s what makes it interesting: Thinking Machines isn’t claiming Inkling is the best model available. Their own blog post says it explicitly. That’s not modesty. It’s strategy.

The Specs

• 975B total parameters, 41B active — Mixture-of-Experts architecture, so you’re not paying for all 975B on every token
• 1M token context window — large enough to hold substantial codebases or document sets
• 45 trillion training tokens — text, images, audio, and video
• Native multimodal reasoning — processes text, images, and audio without bolted-on adapters
• Controllable thinking effort — dial it up for hard problems, dial it down for speed
• Inkling-Small preview — 12B active parameters, same recipe, lower cost

On benchmarks, Inkling holds its own. It ranks near Claude Opus 4.6 on Design Arena’s agentic web dev leaderboard. On Terminal Bench 2.1, it matches Nvidia’s Nemotron 3 Ultra at roughly a third of the tokens. Competitive, but not chart-topping.

Why “Not the Best” Is the Point

Every other major AI lab — OpenAI, Anthropic, Google — builds general-purpose models and sells access to them. The model is the product. You use it as-is.

Thinking Machines is making a different bet: the model is the starting material.

Their platform, Tinker, lets organizations fine-tune Inkling on their own data. The thesis is that a good foundation model, customized with domain expertise, will outperform a stronger general model that knows nothing about your business.

They have a proof point. In a published case study with Bridgewater Associates — the world’s largest hedge fund — researchers fine-tuned an open-source model on Bridgewater’s financial expertise. The result scored 84.7% on financial reasoning benchmarks, beating top proprietary models, at roughly one-fourteenth the cost to run.

That’s the pitch: take our model, add your knowledge, get something better than what you can rent.

The Bigger Picture

This release lands in the middle of a growing argument about who should own the AI that runs your business.

Microsoft CEO Satya Nadella warned last week that enterprises using proprietary AI models effectively pay twice — once in subscription fees, and again by feeding their business knowledge into prompts and corrections that get absorbed into future model versions. Your expertise trains their next release.

Hugging Face CEO Clem Delangue made a similar case: frontier models for experimentation and high-value tasks, open models for everything in production.

Inkling is built for exactly that split. It’s not trying to replace Claude or GPT as your go-to reasoning engine. It’s trying to be the model your organization makes its own.

What This Means Practically

For most professionals and small teams, Inkling is an API model, not a local one. At 975B parameters, you need specialized GPU infrastructure or a 512GB+ Mac to even attempt running it. But with only 41B active parameters per forward pass, API inference costs should be competitive with the big closed models.

The real impact is downstream. Every major open-weight release pressures pricing across the industry. When a model this capable is available under Apache 2.0, the closed labs have to justify their margins with measurably better performance — not just convenience.

For enterprises with the resources to fine-tune, the calculus is shifting. Why pay per-token for a general model when you can own one that already knows your domain?

The Speed Factor

One number stands out: Thinking Machines says it went from founding to revenue in about nine months. OpenAI took roughly five years. Anthropic took three.

That’s partly because Murati and Weng brought deep technical experience from OpenAI. But it also reflects something about where the field is now. The infrastructure exists. The talent pool is global. The playbook is known. Building a frontier lab is no longer a decade-long moonshot — it’s an aggressive but achievable startup timeline.

That compression benefits everyone. More labs building competitive models means more options, better pricing, and faster iteration for the people using them.

Bottom Line

Inkling isn’t going to dethrone Claude or GPT on the leaderboards tomorrow. That’s not the game Thinking Machines is playing.

They’re betting that the future of AI isn’t one model to rule them all — it’s thousands of specialized models, each fine-tuned by the organization that knows its domain best. Inkling is the starting clay. Tinker is the kiln.

Whether that thesis holds depends on whether fine-tuning actually delivers on the Bridgewater promise at scale, across industries, for teams that don’t have hedge-fund-level ML talent. That’s the open question.

But if you’re watching the AI landscape and only tracking the closed frontier labs, you’re missing the story. The open-weight tier just got a serious new entrant — and it’s run by someone who helped build the model everyone else is chasing.