Mira Murati’s Inkling AI Takes on China’s Open-Source Dominance

TL;DR

• Thinking Machines, the startup founded by ex-OpenAI CTO Mira Murati, dropped Inkling — its first open-weight AI model designed as a Western alternative to Chinese open-source systems.
• Inkling handles multimodal queries and can be downloaded, run, and customized by enterprises and researchers — a direct shot at proprietary lock-in from closed-source giants.
• The release comes more than a year after Murati left OpenAI, signaling her bet on openness over the walled-garden approach that now defines her former employer.
• Questions linger about Inkling’s benchmark performance and licensing terms, and whether it can actually match the best open models shipping out of Beijing.

Murati’s Thinking Machines Bets on Openness

Thinking Machines released Inkling more than a year after Mira Murati founded the startup, marking a sharp departure from the closed-source philosophy that now dominates OpenAI. According to Reuters, the model is “open-weight, meaning users can download, run and customize the underlying systems, unlike proprietary, closed-source models.” That’s a direct challenge to the API-only model that’s become the default for frontier labs.

Inkling is designed to process queries across different media — text, images, and other modalities — while balancing cost against performance. The pitch is clear: enterprises and researchers get a high-performance system they can actually control, deploy on-premises, and tweak for compliance or security without begging a cloud provider for API access. For organizations tired of vendor lock-in, that’s a compelling offer.

But the real story here isn’t just what Inkling does. It’s who’s behind it and why they’re doing it this way.

Why Murati’s Exit from OpenAI Still Echoes

Murati left OpenAI amid escalating debates over openness, safety, and commercialization. Her departure wasn’t quiet — it came during a period when OpenAI pivoted hard toward closed models and tighter partnerships with Microsoft. The company that once published GPT-2 with fanfare now guards its weights like nuclear codes.

And here’s Murati, barely a year later, shipping an open-weight flagship model. The contrast couldn’t be sharper. It’s a statement as much as a product launch — a bet that the future of AI doesn’t belong exclusively to a handful of API gatekeepers. Rising developer demand for customizable systems that can be deployed on-premises for security, compliance, and cost control is real, and Murati’s clearly betting Thinking Machines can capture that market.

I think this move also reflects a broader frustration with how frontier labs have closed ranks. When the most capable models are locked behind paywalls and rate limits, innovation gets bottlenecked. Murati’s lived inside that world. She knows exactly what developers are missing.

Inkling Faces a Crowded — and Skeptical — Field

Here’s the problem: Inkling enters a space that’s already packed. Meta’s Llama line has become the de facto standard for open-weight models in the West, with a massive developer ecosystem and solid performance across benchmarks. Community models like Mistral and various fine-tunes have carved out niches. And then there’s the elephant in the room — Chinese labs have been shipping extremely capable open-source and open-weight models at a pace that’s left Western competitors scrambling.

Early coverage focuses on Inkling’s positioning against Chinese open-source leaders, but questions remain about its benchmark performance, licensing terms, and whether it can match or exceed the best open models from China and US big-tech labs. Those aren’t small questions. Developers won’t adopt Inkling just because Murati’s name is on it — they’ll adopt it if it outperforms Llama, matches DeepSeek, and ships with licensing that actually enables commercial use without gotchas buried in the fine print.

The competitive stakes are straightforward. If Inkling can’t hang with the best open models on standard benchmarks, it becomes a curiosity rather than a tool. If it can, it gives the US and European AI ecosystem a high-profile open-weight option at a time when many cutting-edge non-closed models are coming out of China. That matters for organizations that want strong capabilities without full lock-in to proprietary cloud APIs — or to models trained and controlled by labs in Beijing.

Think of it like this: the open-weight AI landscape right now is a bit like the smartphone wars circa 2010. You’ve got the dominant closed ecosystem (Apple/OpenAI), a scrappy open alternative with massive reach (Android/Llama), and a flood of international competitors undercutting on price and features (Chinese OEMs/Chinese AI labs). Inkling is trying to be the Pixel — a Western-backed, high-quality open option that appeals to people who want openness but don’t trust the cheapest option on the market.

But here’s what I keep coming back to: we don’t have the benchmarks yet. We don’t know how Inkling scores on MMLU, HumanEval, or multimodal tasks compared to Llama 3.2 or the latest from DeepSeek. Without those numbers, this is a launch built on reputation and positioning, not proof.

What This Signals About the Open vs. Closed Debate

Murati’s decision to go open-weight with her first flagship model is a vote of no confidence in the closed-source trajectory that’s dominated the last two years. OpenAI, Anthropic, and Google have all tightened access to their best models, arguing that safety and misuse risks demand control. Murati’s counter-argument — implicit in Inkling’s release — is that openness doesn’t have to mean chaos.

The timing matters. Developers are increasingly frustrated with API rate limits, pricing unpredictability, and the inability to fine-tune or deploy models in sensitive environments. Regulated industries like healthcare and finance can’t always send data to a third-party API, no matter how good the model is. Inkling targets that pain point directly.

And it’s not just enterprises. Researchers need access to model weights to study behavior, test safety interventions, and build on top of existing systems. Closed models make that impossible. If Inkling delivers on performance, it could become a go-to for academic labs that want cutting-edge capabilities without the black-box problem.

But the elephant in the room is geopolitics. Chinese labs have been flooding the zone with open models, and Western policymakers are nervous. If the best open-weight models all come from Beijing, that creates dependencies and influence that make Washington uncomfortable. A credible Western alternative — especially one backed by someone with Murati’s pedigree — could ease some of that anxiety.

What to Watch as Inkling Rolls Out

The first thing to watch is benchmarks. Thinking Machines needs to publish detailed performance comparisons across standard evals — MMLU, GSM8K, HumanEval, and multimodal tasks. If Inkling trails Llama 3.2 by a meaningful margin, adoption will stall. If it matches or beats Meta’s latest, this becomes a serious contender overnight.

Licensing terms are the second critical piece. “Open-weight” can mean a lot of things, and the devil is always in the fine print. Can enterprises use Inkling for commercial products without restrictions? Can researchers publish papers based on fine-tuned versions? The more permissive the license, the faster adoption will scale. Any weird carve-outs or usage restrictions will kill momentum before it starts.

Finally, watch how Chinese labs respond. If Inkling gains traction as a Western alternative, expect DeepSeek, Alibaba, and others to accelerate their own open-model releases. The competition here isn’t just technical — it’s geopolitical. Murati’s bet is that there’s demand for a high-quality open-weight model that doesn’t come with the baggage of being trained and controlled in Beijing. Whether that demand is real, and whether Inkling can deliver on it, will define Thinking Machines’ trajectory over the next year.

Source: Reuters