Thinking Machines Inkling: Murati's First Open Model

Thinking Machines Lab released Inkling on July 15, 2026, its first public model and the first thing the Mira Murati startup has shipped since it raised one of the largest seed rounds in the industry's history. Reuters, TechCrunch, WIRED, and Bloomberg all covered it the same day. Inkling is a 975-billion-parameter open-weight model that handles text, audio, video, and code, and the company is releasing the weights rather than gating it behind an API-only endpoint.

The number that isn't in the launch: a benchmark ranking. Thinking Machines didn't lead with "Inkling beats GPT-5.6 on SWE-bench" or a top spot on any agentic leaderboard. That absence is deliberate, and it's the most telling thing about the release.

975B weights in the open, with the ranking left blank

Open weights at this size is the spec that matters. 975B parameters puts Inkling in the same weight class as the biggest Chinese open releases of the past few months: Meituan's LongCat line, Z.ai's GLM family, and the perpetually-updated DeepSeek models. Whether Inkling is dense or a sparse mixture-of-experts design matters enormously for who can actually run it, and the launch coverage doesn't nail that down cleanly.

At 975B total parameters, an MoE architecture with a much smaller active count would be the only way most teams touch it on their own hardware. So treat the raw parameter figure as a marketing number until the model card spells out active parameters and the memory footprint. That detail decides whether "open weights" means anything to a company without a GPU cluster.

What Thinking Machines is plainly not doing is competing on a single scalar. Murati's public framing since the company launched has been that one giant closed model, tuned by one lab for everyone, is the wrong default. Inkling is the argument made concrete. Ship the weights, let customers adapt them, and win on how well the thing bends rather than where it lands on a chart the day it drops.

Tinker is the product; Inkling is the thing you point it at

Read the release through Tinker and it makes more sense. Tinker is Thinking Machines' fine-tuning platform, the API the company has been building toward, and it's the reason an open-weight release is a business move rather than a giveaway. A model you can download is a model you can fine-tune, and a fine-tuning platform needs a flagship model worth fine-tuning.

My read: Inkling exists to give Tinker something native to work on. If your pitch is customization over one-size-fits-all, you can't ask customers to bring a closed API you don't control. You need weights, ideally your own, that a paying customer can specialize on their data and keep. Inkling supplies that. The open release and the commercial platform are two ends of the same strategy.

This is a different shape from how OpenAI or Anthropic monetize. Those labs sell inference on models you never hold. Thinking Machines is betting there's a real market for teams that want to own an adapted model outright, run it where they choose, and not rent intelligence by the token forever.

Multimodal on the input side, but check what the weights cover

Inkling takes text, audio, video, and code. Four input modalities in one open release is genuinely uncommon at this scale; most open models that go multimodal stop at images. Whether it generates audio and video or only understands them is the question the spec sheet has to answer, and "multimodal" in a press release routinely means input-only comprehension.

If the released weights include the full audio and video stack rather than a text-only checkpoint with the multimodal parts held back, that's a real gift to anyone building on top. Labs frequently open a partial model and keep the interesting encoders proprietary. Until the Hugging Face card is up and people have loaded it, assume nothing about coverage.

The pitch is "not one-size-fits-all," aimed straight at closed labs

Thinking Machines is positioning Inkling as a Western open-weight option for enterprises that want to customize, at a moment when the strongest open models mostly ship from China. DeepSeek, GLM, LongCat, Qwen: the open frontier has had a distinctly Chinese center of gravity for a year. A US lab founded by OpenAI's former CTO putting a 975B multimodal model under an open license is a direct answer to companies that want open weights but have procurement or policy reasons to avoid Chinese releases.

That's a real gap in the market. Plenty of regulated buyers, defense-adjacent firms, and enterprises with China-sourcing restrictions have wanted a capable open model they can legally deploy and adapt. If Inkling's license is genuinely permissive and the weights perform, it slots into a space that Llama has partly owned and that few others have contested at the frontier.

The catch is that "Western alternative" only sells if the model is good. Enterprises don't adopt a worse model for its passport. The absence of benchmark claims cuts both ways here: it signals confidence in customization, and it also means nobody yet knows how the base model stacks up against Llama, DeepSeek, or a fine-tuned Claude Sonnet on the tasks buyers actually care about.

Where Inkling sits against the open field

Here's the rough landscape as of the July 15 launch, from public specs and coverage. Treat capability rows as claims, not verified results.

What Inkling brings that most of these don't is the platform pairing. Llama has the largest deployment base and the deepest tooling, but Meta doesn't sell you a managed way to fine-tune it as a first-party product. The Chinese models compete hard on price and raw capability. Thinking Machines is trying to win on the full customization loop: download, adapt on Tinker, deploy your version. Whether that loop is smooth enough to matter is the thing to watch over the next few weeks.

What Murati hasn't shown yet

The launch leaves the load-bearing details open. Dense or MoE, and how many parameters are active per token. The exact license terms and whether they're truly permissive or carry usage restrictions. Real evaluation numbers on standard suites so buyers can compare it to Llama and DeepSeek without guessing. Whether the multimodal weights are fully released or a text-first checkpoint. And what Tinker actually costs to fine-tune and serve at production scale.

None of that undercuts the significance of the drop. A well-funded US lab led by a former OpenAI CTO putting a frontier-scale multimodal model under open weights is a serious move, and the customization-first framing is a genuine bet against the closed-API default that OpenAI and Anthropic have built their businesses on. It just means the important part hasn't been settled by the announcement. It'll be settled by whoever loads the weights this week and reports back on whether the model bends the way Thinking Machines says it does.