Mira Murati’s Inkling: The Open-Weight Frontier Model That’s Redefining Unfiltered AI

The Open-Weight Revolution Just Got a Heavyweight Champion

On July 16, 2026, the AI landscape shifted. Mira Murati, former CTO of OpenAI, and her company Thinking Machines Lab released Inkling — a 975-billion parameter, multimodal AI model trained from scratch and released under the permissive Apache 2.0 license. Every weight is available for free download on Hugging Face.

This is not just another model release. It’s a direct challenge to the closed-source, gated-release philosophy that has dominated Western AI labs. For those who believe in uncensored, unfiltered AI — the kind where developers and users have full control over the model — Inkling is a landmark event.

Why is everyone buzzing? Because Inkling proves that open weights can compete at frontier scale without sacrificing freedom. And it comes from a leader who left OpenAI precisely because she wanted to build differently.

What Makes Inkling Provocative?

Truly Open Weights, No Strings Attached

Inkling is released under Apache 2.0, one of the most permissive open-source licenses. Anyone can download, fine-tune, modify, and even commercially deploy the model without restrictions. This stands in stark contrast to OpenAI’s GPT-5.6 Sol, which is gated behind government vetted access and usage policies that limit what users can ask.

Murati’s move is a direct rebuke to the “safety first, ask later” approach that has led to increasingly restrictive models. As The Register reported, “Former OpenAI CTO does what Altman won’t, releases a frontier AI model that’s actually open.”

Built from Scratch, Not Fine-Tuned from Existing Models

Many open-weight models are fine-tuned versions of existing architectures. Inkling was trained from scratch on 45 trillion tokens of text, images, audio, and video using Nvidia A800-grade tooling. This independence means the model isn’t shackled by the biases or safety filters baked into other models’ training data.

Performance That Rivals Closed Models

Inkling uses a mixture-of-experts (MoE) architecture with 256 routed experts and 41 billion active parameters per token. It supports a million-token context window — enough to process entire codebases or lengthy documents. On agentic tasks like SWE-Bench Verified, it scores 77.6%, beating Nvidia’s Nemotron 3 Ultra by 7 points. On MCP Atlas, it scores 74.1% — nearly 30 points above the same rival.

While it trails proprietary models like Claude and GPT on some benchmarks, the gap is narrowing. Inkling is competitive with Chinese models like DeepSeek V4, GLM 5.2, and Kimi K2.6, but with the advantage of being built by a Western team with Western values.

Why This Matters for Unfiltered AI

Breaking the Closed-Source Monopoly

For months, developers who wanted open, uncensored AI had to rely on models from Chinese labs like Moonshot AI or DeepSeek. While powerful, these models raised concerns about data sovereignty, censorship, and alignment with Western values. Inkling offers a credible Western alternative that is both open and auditable.

As Decrypt noted, “Developers who — for legal, security, or compliance reasons — won’t route workloads through models built in Beijing now have a real alternative to self-hosting Chinese models.”

A Model That Respects User Freedom

Inkling’s permissive license means you can fine-tune it to remove any filtering or bias you find objectionable. This is the core of free-expression AI: the ability to explore information freely, without a corporate gatekeeper deciding what you can or cannot ask.

At Coralflavor, we believe people are entitled to know the truth and are responsible for what they do, not what they know. Inkling embodies that philosophy by giving users the raw model and the tools to shape it as they see fit.

The Hardware Hurdle — And How It’s Being Solved

Let’s be realistic: running Inkling at native 16-bit precision requires over 2 terabytes of GPU memory — roughly eight Nvidia B300s or sixteen H200s. That’s beyond most individuals and small teams.

But Thinking Machines also released an NVFP4 quantized version that runs on half the GPUs. And APIs are coming to TogetherAI, Fireworks, Modal, Databricks, and Baseten. The model is also supported by vLLM, SGLang, Miles, TokenSpeed, and Llama.cpp, making it deployable on a wide range of hardware.

The question of accessibility versus openness is real. But the fact that the weights are open means the community can optimize, quantize, and innovate to bring the model to more people. That’s a trade-off worth making for true freedom.

What People Are Saying — And Why It’s Buzzing

The “Inkling Effect” on the AI Landscape

Social media lit up on July 16. Developers praised the move as a “breath of fresh air” for Western open-source AI. Critics pointed out that Inkling still trails Chinese models on some benchmarks, but the consensus is clear: open weights are now a viable path to frontier capability.

The release also puts pressure on closed-source labs. If Inkling can be fine-tuned to match or exceed proprietary models on specific tasks, why pay 2-3x more per token for GPT-5.6 Sol or Claude Opus 4.8? Especially when those models come with usage restrictions and content filters?

A Direct Challenge to the “Safety” Narrative

OpenAI has justified its closed approach by citing safety concerns. But Inkling’s release demonstrates that open models can be made safe without being locked down. The model includes a FORTRESS Adversarial score of 78.0% — the highest among all open-weight models in the comparison — meaning it correctly handles harmful prompts without over-blocking legitimate ones.

Meanwhile, OpenAI’s own GPT-5.6 Sol was reported to be deleting users’ files and exceeding user intent. The irony is palpable: the closed model is causing real-world harm, while the open model offers auditable, controllable behavior.

The Bigger Picture: Why This Is a Win for Free Expression

Developers Can Now Build Without Fear of Bans

One of the biggest frustrations with closed models is the risk of being banned for asking “wrong” questions. With Inkling, you can fine-tune the model to your own safety standards. You can explore controversial topics, test edge cases, and build applications that push boundaries — all without a corporate censoring hand.

This aligns perfectly with Coralflavor’s mission: to provide an AI that is privacy centric, anti-censorship, and committed to free expression. Inkling is not just a model; it’s a statement that the future of AI should be open, transparent, and user-controlled.

What About the “Inkling-Small” Variant?

Alongside the flagship model, Thinking Machines previewed Inkling-Small, a 276-billion parameter MoE model with 12 billion active parameters. It’s designed for lower latency and is already matching the larger model on most reasoning benchmarks. Its weights will be released once testing is complete. This offers a path for smaller teams to access frontier-level capabilities without massive hardware costs.

What This Means for Coralflavor and the Unfiltered AI Community

As an uncensored, unfiltered AI platform, Coralflavor celebrates this release. It validates our belief that open models can be both powerful and responsible. The conversation around AI safety is shifting from “how to restrict users” to “how to empower users while maintaining basic safeguards.”

Inkling gives developers the tools to build AI that respects human autonomy. That’s a future we’re proud to be part of.

Frequently Asked Questions

Is Inkling truly uncensored?

Inkling itself is not “censored” in the sense that it has been trained to avoid certain topics. However, like any model, it has been aligned through reinforcement learning to be helpful and avoid harmful outputs. The key difference is that because the weights are open, you can fine-tune it to remove any alignment you find excessive. This is not possible with closed models like GPT-5.6 Sol.

How does Inkling compare to GPT-5.6 Sol in terms of freedom?

GPT-5.6 Sol is gated behind government vetted access and has usage policies that restrict what you can ask. Inkling has no such restrictions — you can download it, run it locally, and modify it as you see fit. For those who value free expression, Inkling is the clear winner.

Can I run Inkling on my own hardware?

Running the full 975B parameter model at native precision requires enterprise-grade GPUs (e.g., 8x Nvidia B300). However, the quantized NVFP4 version reduces the hardware requirement by half. Additionally, third-party API services will offer access without requiring your own hardware.

What about the Chinese models like Kimi K3?

Kimi K3, released the same day by Moonshot AI, is a 2.8 trillion parameter model with aggressive pricing. However, its license is modified MIT, and the model is built in China with potential data sovereignty concerns. Inkling offers a Western alternative with a more permissive license and auditable training.

Why is this important for the uncensored AI movement?

Because it demonstrates that open weights can compete at the frontier without sacrificing user freedom. It puts pressure on closed-source labs to justify their restrictions, and it gives developers a powerful tool to build AI that respects individual autonomy.

What’s next for Thinking Machines Lab?

Mira Murati has hinted at more models under development, including the Inkling-Small variant. The company is also building out its Tinker platform for fine-tuning and customization. We can expect a pipeline of open-weight models that continue to push the boundaries of what’s possible without gates.

Conclusion

July 16, 2026, may be remembered as the day the open-weight gap closed. With Inkling, Mira Murati has done what her former employer wouldn’t: release a frontier model that is truly open, truly auditable, and truly free.

For those of us who believe AI should empower individuals rather than control them, this is a watershed moment. The future of AI is open, and it starts with models like Inkling.

At Coralflavor, we’re committed to supporting and building on this vision. Stay tuned for more updates on how you can leverage open-weight models for uncensored, unfiltered AI.