Mira Murati's 975B-parameter open AI model challenges OpenAI

Key Takeaways

• Thinking Machines Lab released Inkling, a 975 billion parameter open weights model under Apache 2.0 license
• The model requires 2TB+ of GPU memory but offers quantized versions for smaller deployments
• Inkling competes with Chinese models like DeepSeek V4 while offering full commercial freedom for fine-tuning

Mira Murati, who left her CTO role at OpenAI in September 2024, just shipped what her former employer refuses to build: a frontier-class open weights AI model. Thinking Machines Lab released Inkling on Wednesday, a 975 billion parameter model available under Apache 2.0. For enterprises tired of API lock-in and Chinese model dependencies, this is the first serious American alternative.

What makes Inkling different from existing open models?

Size, mostly. At 975 billion parameters, Inkling dwarfs Meta's Llama 3.1, which tops out at 70 billion. It sits in the same weight class as DeepSeek V4, GLM 5.2, and Kimi K2.6. The difference: you can run it on your own infrastructure, fine-tune it for proprietary workflows, and deploy it without routing data through Beijing.

The hardware requirements are substantial. Running Inkling at native 16-bit precision demands more than two terabytes of GPU memory. That translates to roughly eight Nvidia B300 accelerators or sixteen H200s. For organizations with smaller GPU clusters, Thinking Machines offers an NVFP4 quantized version that halves the hardware requirement.

Thinking Machines trained the model from scratch on 45 trillion tokens of text, images, audio, and video using Nvidia GB300 NVL72 systems. The architecture borrows from DeepSeek-V3's mixture of experts approach, with 256 routed experts and two shared ones. Each token generation activates six experts, totaling about 41 billion active parameters. Despite the massive total size, inference speed should match DeepSeek V4 on equivalent hardware.

How does Inkling perform against closed models?

Thinking Machines claims Inkling competes with Chinese open models across various workloads, though the company's own benchmarks show it trailing proprietary systems from Anthropic and OpenAI. Benchmark claims in AI deserve skepticism. Gaming these evaluations isn't difficult, and every model maker cherry-picks favorable results.

One concrete efficiency claim: Inkling reportedly matches Nvidia's Nemotron 3 Ultra on Terminal Bench 2.1 using roughly a third of the reasoning tokens. That matters for cost. Reasoning models bill for every token they generate while "thinking," so fewer thinking tokens means lower inference costs at scale.

The model supports a million-token context window. For enterprise use cases involving large codebases or document retrieval, this should help with the classic needle-in-haystack problem.

What does Apache 2.0 licensing mean for enterprises?

Apache 2.0 is about as permissive as open source gets. You can modify, distribute, and commercialize derivatives without sharing your changes. Compare this to OpenAI's GPT-4o or Anthropic's Claude, where you're permanently dependent on their API, their pricing, and their content policies.

The practical value: your legal and compliance teams can audit the model. Your ML engineers can fine-tune it for domain-specific tasks. Your data never leaves infrastructure you control. For regulated industries, healthcare, and defense contractors, this eliminates entire categories of vendor risk.

Thinking Machines' Tinker platform provides tooling for customization. The company claims Inkling can write its own fine-tuning scripts, evaluate its abilities, and teach itself new skills. Third-party API access is coming through TogetherAI, Fireworks, Modal, Databricks, and Baseten. If you prefer self-hosting, weights are available on Hugging Face with support for vLLM, SGLang, Miles, TokenSpeed, and Llama.cpp.

The strategic context: why Murati built this

Murati left OpenAI during the leadership turmoil following Sam Altman's brief ouster. Her departure was part of a broader executive exodus. She founded Thinking Machines Lab in early 2025, positioning it as research-focused and, crucially, committed to open release.

The timing matters. American enterprises wanting frontier-class open weights had few options outside Chinese model houses. Meta's Llama models are open but significantly smaller. OpenAI abandoned its founding "open" mission years ago. Murati is filling a gap that her former employer created.

Thinking Machines is also previewing Inkling-Small, a 276 billion parameter MoE model with 12 billion active parameters. This variant prioritizes latency over throughput for applications where response time matters more than raw capability. Weights will release after testing completes.

Frequently Asked Questions

How much GPU memory does Inkling require?

The full 16-bit model needs over 2TB of GPU memory, roughly eight Nvidia B300 or sixteen H200 accelerators. A quantized NVFP4 version runs on half that hardware.

Can I use Inkling for commercial products?

Yes. The Apache 2.0 license permits modification, redistribution, and commercial use without royalties or sharing requirements.

How does Inkling compare to DeepSeek V4?

Both models use mixture-of-experts architectures and have similar parameter counts. Inkling claims competitive performance with the added benefit of Apache 2.0 licensing and U.S.-based development.

Where can I download Inkling?

Model weights are available on Hugging Face. API access is available through Thinking Machines' Tinker platform, with third-party providers like TogetherAI and Fireworks coming soon.

Source: www.theregister.com