Thinking Machines amps up its bet against one-size-fits-all AI with its first open model, Inkling

Mira Murati's year-and-a-half stealth build has a name now: Inkling. The former OpenAI CTO's startup, Thinking Machines Lab, dropped its first in-house model Wednesday morning. It's open-weight. That means anyone can download the thing, crack it open, and rewrite its guts. OpenAI, Anthropic, and Google have never let you do that with their flagships.

The architecture is a mixture of experts — 975 billion total parameters, but only 41 billion fire for any single request. That sparse activation pattern isn't novel. It's how you make gargantuan models tractable. Inkling was fed 45 trillion tokens spanning text, images, audio, and video. The company says it reasons natively across all four modalities. For now, though, it only speaks text. Code, styled artifacts, structured data. No voice. No video out.

Here's the real story: Thinking Machines isn't pretending Inkling wears a crown. Their own blog post states flatly that it's "not the strongest overall model available today, open or closed." That honesty is rare. It's also a tell. They're not playing the leaderboard game. They're playing a different game entirely.

The bet is that one-size-fits-all models — ChatGPT, Claude, Gemini — hit a ceiling because they're frozen at the factory. Central training locks in generic knowledge. But the expertise that actually moves money inside organizations lives in the heads of specific people, in specific workflows, in specific data that never sees the public internet. Murati's argument, laid out in a manifesto posted last week, is that AI shaped by the organization itself will outperform AI shaped by a lab in San Francisco.

Inkling is the test vehicle. It's designed to flag uncertainty instead of hallucinating. It lets operators dial "thinking effort" up or down, trading latency for depth. On one coding benchmark, the company claims Inkling matches Nvidia's Nemotron 3 Ultra while burning a third of the tokens. If that holds under independent scrutiny, it's a meaningful efficiency signal. But it's one benchmark. One vendor claim. No third-party replication yet.

The enterprise pitch is explicit: Inkling isn't a product. It's a starting point. Customers are meant to fine-tune it themselves via Tinker, Thinking Machines' customization platform. That shifts the heavy lifting — and the liability — onto the buyer. Fine-tuning a 975-billion-parameter mixture-of-experts model demands serious machine-learning engineering. Most enterprises don't have that bench depth. They have API keys and prompt engineers. Thinking Machines knows this. They're betting the market will mature fast enough to catch up.

Safety follows the same logic. The startup washes its hands of what happens after customization. If your fine-tuned Inkling leaks PII or advises fraud, that's on you. That's a defensible stance for a platform play. It's a dangerous one for a brand trying to win conservative IT buyers.

Contrast the incumbents. OpenAI built ChatGPT as a consumer chatbot first, then bolted on agentic features. Anthropic did the same with Claude. Google with Gemini. Their DNA is general-purpose conversation. Autonomous action came later, layered on top. Thinking Machines started with a different premise: interaction models that listen, speak, interrupt — models that behave like colleagues rather than oracles. The May research preview showed that lineage. Inkling inherits it.

The modal gap matters. Native multimodal reasoning with text-only output is a half-delivered promise. Enterprises drowning in video, audio, and image data need models that emit those formats, not just ingest them. Until Inkling generates across modalities, it's a thinker that can't show its work in the mediums where the work lives.

Pricing remains opaque. No public API tiers. No self-serve sign-up. The go-to-market motion suggests high-touch enterprise sales — pilots, proof-of-concepts, professional services. That's where the revenue lives. But it also caps adoption speed. Developers can't kick tires on a Saturday afternoon.

The open-weight release is the sharpest edge. Meta's Llama line proved that open weights catalyze an ecosystem: quantizers, fine-tuners, tool builders, researchers. Thinking Machines wants that gravity. But Llama had first-mover advantage and Meta's distribution muscle. Inkling enters a crowded field. Nvidia's Nemotron 3 Ultra already sits in the open-weight lane with serious benchmark momentum. Alibaba's Qwen family owns the Chinese enterprise flank. Mistral plays the European sovereign card. Inkling needs more than architecture to carve territory.

Murati's pedigree buys credibility. OpenAI's CTO doesn't leave to build a also-ran. But credibility isn't product-market fit. The enterprise buyers Thinking Machines courts — banks, healthcare, manufacturing, legal — move slow. They need compliance artifacts, audit trails, vendor stability. A year-old startup with one model and a customization platform that demands PhD-level ML ops is a hard sell to procurement.

The skepticism is warranted on multiple fronts. Token efficiency claims need reproduction. Multimodal output is absent. The fine-tuning burden is real and steep. Safety ownership transfers to the least equipped party. And the core thesis — that organizations will effectively distill their proprietary expertise into model weights — assumes a capability maturity that mostly doesn't exist.

But the directional bet is sound. The era of monolithic closed models serving every master equally is ending. Specialization wins. The winners will be the platforms that make specialization tractable for organizations that don't employ hundreds of researchers. Tinker is the real product. Inkling is just its first demonstration.

Whether Thinking Machines can turn that demonstration into a defensible business before the incumbents copy the playbook — that's the only question that matters.