The Creator of MLX Just Left Apple - And He's Not the First

Awni Hannun, the researcher who co-created MLX - Apple's open-source machine learning framework for Apple Silicon - announced on Friday that he has left the company.

"Today is my last day at Apple. Building MLX with our amazing team and community has been an absolute pleasure. It's still early days for AI on Apple silicon. Apple makes the best consumer hardware on the planet. There's so much potential for it to be the leading platform for AI."

The farewell was gracious. He named the team members carrying MLX forward - Angelos Katharopoulos, Cheng Zhao, Jagrit Digani, and others - and expressed confidence in the project's future. He did not say where he's going next.

What Hannun didn't address is the question that matters most: why the person who built the most important open-source ML project Apple has ever shipped decided it was time to walk away - and whether MLX can survive the broader collapse of Apple's AI research bench.

Who Awni Hannun Is

Hannun isn't a mid-career engineer who happened to start a side project. He's one of the more accomplished ML researchers of his generation.

The research record

He holds a PhD in Computer Science from Stanford, where he was advised by Andrew Ng. At Baidu's Silicon Valley AI Lab, he co-led the Deep Speech project - an end-to-end deep learning speech recognition system that, in 2014, outperformed all existing approaches on the Switchboard benchmark. Deep Speech 2 extended the architecture to both English and Mandarin and was published at ICML 2016.

At Facebook AI Research (FAIR), he worked on speech recognition, privacy-preserving ML, and framework development. His paper on certified data removal from machine learning models won the Best Paper award at UAI 2021. His work with Andrew Ng on cardiac arrhythmia detection was published in Nature Medicine - a deep neural network that achieved an average F1 score of 0.837, exceeding that of average board-certified cardiologists (0.780).

Across 37 publications spanning Nature Medicine, ICML, NeurIPS, and Interspeech, Hannun built a reputation as someone who ships research that works in the real world.

What he built at Apple

Hannun co-created MLX and served as its most visible champion. The framework, released in December 2023, gave researchers and developers a native way to run machine learning workloads on Apple Silicon - exploiting the unified memory architecture that makes M-series chips unique. MLX's Python API mirrors NumPy, its higher-level packages mirror PyTorch, and it ships under the MIT license.

MLX enabled things Apple's proprietary CoreML never could. Researchers ran DeepSeek R1 680B across three M2 Ultras. Fine-tuning with LoRA became possible on consumer MacBooks. LM Studio integrated MLX as a first-class inference engine. The mlx-community on HuggingFace grew to over 1,000 converted model weights. In July 2025, the team began adding CUDA support - led by Cheng Zhao, the creator of Electron - transforming MLX from Apple-exclusive to cross-platform.

As recently as mid-February, Hannun was tweeting about Qwen 3.5 running on MLX and building voxmlx, a speech recognition package he wrote completely with Claude Code. "I did not write a single line of code," he noted.

Two weeks later, he left.

MLX turned Apple Silicon into a first-class ML research platform - 24,100 GitHub stars and an ecosystem that includes LM Studio, HuggingFace, and distributed inference over Thunderbolt 5.

The Exodus

Hannun's departure does not exist in isolation. It's the latest in what analysts have called a "crisis of confidence" in Apple's AI future - a sustained hemorrhage of senior ML talent that has gutted the company's research bench over the past twelve months.

The timeline

July 2025 - Ruoming Pang, head of Apple Foundation Models, leaves for Meta with a package reportedly worth over $200 million. Pang led approximately 80 engineers and had spent 15 years at Google before joining Apple in 2021. On July 12, he made a "final appeal" to Apple's Software Engineering division requesting permission to publicly share research results. The request was denied. He resigned the same month.

Mid-2025 - The MLX team itself threatened to quit. Apple scrambled with emergency counteroffers to retain the core developers. Analytics India Magazine reported at the time: "The mutiny was averted, but the fact that it happened at all is proof that morale in the iPhone maker's camp is shaky and confidence in leadership is eroding."

December 2025 - John Giannandrea, SVP of Machine Learning and AI Strategy since 2018, announced his retirement. Internal reports suggested CEO Tim Cook had lost confidence in Giannandrea's ability to deliver on Siri improvements. The personalized Siri update - originally planned for late 2025 - slipped to iOS 26.4.

January 2026 - Four more AI researchers departed in a single week: Yinfei Yang (multimodal lead, 4,000+ citations) left to start a company. Haoxuan You joined Meta's Superintelligence Labs. Bailin Wang went to Meta. Zirui Wang went to Google DeepMind - where he now helps build models that Apple pays Google to license.

February 2026 - Awni Hannun, the public face of MLX, leaves Apple.

Apple has lost approximately a dozen AI researchers since early 2025, including its head of foundation models, its AI chief, and now the creator of MLX.

Why they're leaving

The departures share common root causes, documented across multiple reports:

Secrecy culture - Apple's SVP of Software Engineering Craig Federighi maintained strict confidentiality requirements that prevented researchers from publishing papers, presenting at conferences, or sharing results until technology shipped in products. One source told KR Asia: "Craig would rather let people say Apple is bad at AI than let them say Apple builds bad AI services." For researchers whose careers depend on publication records, this was untenable.

Compensation gaps - Meta offered Pang over $200 million. Yu Jiahui, a multimodal specialist, was recruited for around $100 million. Apple's traditional compensation structures - which rely heavily on long-vesting stock grants - couldn't match the accelerated packages, signing bonuses, and publication freedoms that Meta and Google were offering.

Strategic confusion - Apple's AI strategy shifted repeatedly. Build in-house? Partner with OpenAI? Pivot to Google Gemini? Each change left researchers uncertain about whether their work would ever ship. As one industry analyst put it: "If you want to work on cutting-edge AI, Apple is not the place."

The Strategic Contradiction

Here is the tension that makes Hannun's departure significant beyond one person leaving one company.

MLX represents a specific philosophy: open-source, on-device, researcher-empowering ML that runs locally on Apple Silicon. It gives developers the tools to run frontier models on their own hardware without cloud dependencies.

Apple's strategic direction has moved in the opposite direction.

In January 2026, Apple and Google announced a multi-year partnership under which Apple's next-generation foundation models will incorporate Google's Gemini models and cloud infrastructure. Apple is reportedly paying Google approximately $1 billion annually. Gemini is becoming the default intelligence layer for Apple's AI products, while the company's on-device model remains a modest ~3B-parameter system handling basic tasks.

Apple now maintains three separate ML frameworks with different purposes:

MLX lives in Apple's Machine Learning Research group - not the product engineering org. It has no official role in Private Cloud Compute or Apple Intelligence's production pipeline. It received two dedicated WWDC 2025 sessions and continues to be actively developed. But its philosophical alignment with Apple's corporate strategy grows weaker with every cloud partnership announcement.

The person who built the bridge between Apple's hardware advantage and the open ML community just walked off the bridge.

What Happens to MLX

Hannun named his successors. The team includes Angelos Katharopoulos (architect of MLX's distributed RDMA backend for Thunderbolt 5 inference), Cheng Zhao (leading CUDA cross-platform support), Jagrit Digani (original co-creator), and Ronan Collobert (former FAIR researcher). The technical bench is real.

But Hannun was more than a coder. He was MLX's public evangelist - the person who tweeted benchmarks, explained design decisions, engaged with the community, and made the project feel like it belonged to developers rather than a corporation. Replacing that role at a company whose default posture is silence will be difficult.

The framework's 24,100 GitHub stars and 209 contributors suggest enough community momentum to survive leadership changes. The CUDA expansion, the LM Studio integration, and the HuggingFace ecosystem all create inertia. But frameworks aren't self-sustaining. They need champions with the authority and willingness to fight internal battles over resources, direction, and visibility. Hannun had that authority. Whether his successors will is an open question.

Hannun's farewell post was warm. He praised the team. He praised the hardware. He expressed confidence in MLX's future. He said nothing about Apple's leadership, its strategy, or his reasons for leaving.

That silence, at a company famous for enforcing it, says more than any statement could.

Sources:

• Awni Hannun's departure announcement - X
• Trapped by silence: How Apple lost its top AI talent to Meta - KR Asia
• Apple's top LLM researcher quits, MLX team threatened to leave - Analytics India Magazine
• Apple AI Chief John Giannandrea Retiring After Siri Delays - MacRumors
• Apple loses more AI researchers and a Siri executive - 9to5Mac
• MLX - GitHub
• Awni Hannun publications
• Apple's machine learning framework is getting support for NVIDIA GPUs - 9to5Mac