Inkling Is Thinking Machines’ Big Bet on an Imperfect Open Model

Thinking Machines Lab released Inkling on Wednesday, a 975 billion parameter open-weights AI model, and led with an admission: it isn’t the best one available. The startup, founded by former OpenAI chief technology officer Mira Murati, built what is now the largest open-weights model to come out of the United States. It ships free to download, inspect and fine-tune under an Apache 2.0 license.

The wager is straightforward. Thinking Machines is betting that enterprises will pay to reshape an admittedly average model they can own outright, even as closed rivals and Chinese open-weight labs keep pulling further ahead on raw intelligence. Independent testing already shows one cost of that bet. Artificial Analysis, an independent benchmarking platform, measured a 63% hallucination rate on the same model it otherwise ranks as the strongest open-weights system built in America.

A 66-Layer Model Built to Be Taken Apart

Inkling is a sparse Mixture-of-Experts transformer, 66 layers deep. The checkpoint holds 975 billion total parameters, but only about 41 billion activate for any given token. Each layer routes work through 256 routed experts and two shared experts, with six routed experts firing per token.

The model supports a context window of up to one million tokens and was pretrained on 45 trillion tokens of text, images, audio and video. Thinking Machines trained it entirely on Nvidia’s GB300 NVL72 systems, part of a March deal to deploy a gigawatt of Vera Rubin computing capacity. Building it took about nine months and roughly 200 employees.

Inkling’s routing design borrows ideas popularized by DeepSeek-V3, the Chinese model it was built to compete against, but Thinking Machines trained it from scratch instead of fine-tuning someone else’s weights. The company confirmed the release on X, writing that the full weights are available for fine-tuning on Tinker starting immediately. In its own announcement, Thinking Machines admits Inkling is not the strongest model available today, open or closed.

That openness has limits. The license field reads Apache 2.0, but a separate terms of use narrows how the model can actually be deployed, according to one detailed technical review of the release documents. Thinking Machines’ vision of AI shaped differently by every organization only works once hardware, contracts and technical skill line up.

Murati Is Betting Against the Playbook She Helped Write

Axios framed the wager plainly: Thinking Machines is making a different bet than most AI labs, one that says enterprises care less about the smartest general-purpose model than about one they can make their own. The company is emphasizing speed, too. OpenAI took roughly five years to bring its technology to market and show revenue. Anthropic took about three. Thinking Machines says it did the same in about nine months.

Behind Inkling sits a broader idea Murati has been describing for months, about AI built for real-time back and forth instead of stop-and-wait chat.

Our interactions with each other are very rich.

Murati told Bloomberg last month. She has argued that meaning lives in silence, hesitation and interruption, texture a typical chatbot throws away, and that Inkling is designed to serve as the background reasoning model for the interaction models system Thinking Machines previewed in May.

Thinking Machines is not trying to make money from Inkling directly. Revenue comes through Tinker, its fine-tuning platform, which it already sells to customers including hedge fund Bridgewater Associates. The two companies fine-tuned an open model on Bridgewater’s financial expertise and reported a score of 84.7% on financial reasoning tests, beating proprietary rivals while costing roughly a fourteenth as much to run. That figure comes from the two companies’ own evaluation, not an independent one.

The Calibration Push That Still Hallucinates

Thinking Machines spent real engineering effort on a quality most labs treat as an afterthought: teaching Inkling to know what it doesn’t know. The company trained it with reinforcement learning against proper scoring rules on a large corpus of resolved real-world questions, and built a claims grader that checks factual statements with agentic web search instead of trusting the model’s own memory.

The forecasting numbers back that up. On ForecastBench without search access, Inkling scores 61.1 on the Brier Index, matching Google’s Gemini 3.1 Pro and beating Anthropic’s Claude Opus 4.8 at 54.6. On SimpleQA Verified, a pure factuality test, it scores 43.9%.

None of that stopped Artificial Analysis from clocking that 63% hallucination rate, alongside weaker factual accuracy than the benchmark charts alone suggest. Independent reviewers have also flagged that Thinking Machines’ own release materials do not fully agree with each other.

The launch page and model card reported 73.5% on MMMU-Pro, while the commit-pinned Hugging Face README listed 73.3%.

• Thinking Machines says Inkling matches Nvidia’s Nemotron 3 Ultra on Terminal Bench 2.1 using roughly a third of the tokens.
• Artificial Analysis ranks Inkling the strongest open-weights model from a US lab, while separately reporting that 63% hallucination rate.
• MarkTechPost’s reading of the same benchmark chart shows Inkling trailing GLM 5.2 by 18.9 points on that identical Terminal Bench 2.1 test.

Thinking Machines flagged occasional compliance with harmful role-play prompts in its own model card, and recommends layering moderation tools like Llama Guard around any deployment instead of trusting the model’s refusals alone. A widely discussed academic paper on restricting access to AI models undermining the safety it protects argues the open-versus-closed safety debate is more complicated than either side admits, a framing that fits Inkling’s mixed report card.

How Much Hardware Does Running Inkling Actually Take?

Running Inkling at full precision takes more than two terabytes of combined GPU memory, roughly eight Nvidia B300 accelerators or sixteen H200s. A quantized NVFP4 version cuts that to around 600 gigabytes, still several high-end GPUs. Hosted access through Tinker and outside partners lowers the barrier, but nobody is fine-tuning this model on a laptop.

• BF16, full precision – needs at least 2 TB of aggregated VRAM, on 8x Nvidia B300 or 16x Nvidia H200 accelerators.
• NVFP4, quantized – drops that requirement to roughly 600 GB of VRAM, running on 4x B300 or 8x H200.
• Tinker hosted fine-tuning – live now with 64K and 256K context options, no local hardware required.
• Third-party inference – rolling out through TogetherAI, Fireworks, Modal, Databricks and Baseten.

Day-zero support already exists in vLLM, SGLang, Hugging Face transformers and Llama.cpp, so hobbyists with heavily quantized local builds can experiment even without a server rack.

Nvidia Profits Whether Inkling Wins or Not

Thinking Machines struck a partnership with Nvidia in March to deploy a gigawatt of Vera Rubin computing capacity, and trained Inkling entirely on Nvidia’s GB300 NVL72 systems. Nvidia is also among the startup’s investors. The two companies have deepened that relationship since.

Developers may or may not adopt Inkling as a finished product. Running it still means buying more Nvidia hardware, whether at two terabytes for full precision or 600 gigabytes quantized. One detailed technical analysis of the release concluded that running the full model requires significant GPU infrastructure, which keeps closed-model APIs more economical for many organizations than self-hosting a 975 billion parameter system.

Thinking Machines’ spending may never need to reach the scale of OpenAI’s or Anthropic’s. Once weights are public, nobody who downloads them owes Thinking Machines anything to run them, unlike the metered access its closed rivals sell. TechCrunch reported that a $50 billion funding round appeared to be coming together for the startup last November before stalling by January, with the company declining to discuss its funding picture since.

Does Inkling Actually Beat China’s Open Models?

Sometimes. Thinking Machines’ own numbers put Inkling ahead of Kimi K2.6 and DeepSeek V4 Flash Max on agentic and financial benchmarks, and its efficiency claim against Nemotron 3 Ultra holds up. GLM 5.2 still posts a clearly higher score on the same coding test Thinking Machines highlights, and DeepSeek’s larger variant beats Inkling on pure coding accuracy.

Inkling scored 77.6% on SWE-Bench Verified, behind DeepSeek V4 Pro and GLM 5.2 but ahead of Nvidia’s Nemotron 3 Ultra. On Terminal Bench 2.1, Inkling posts 63.8%, while GLM 5.2 comes in about 18.9 points higher, near 82.7%, even though Inkling reaches its score using a fraction of the tokens.

Artificial Analysis puts Inkling at 41 on its Intelligence Index, three points above the previous leading US open model, Nemotron 3 Ultra, and well ahead of Google’s Gemma 4 31B at 29 and OpenAI’s gpt-oss-120b at 24. That makes Inkling the strongest open-weights model built by an American lab. It is not the strongest open-weights model in the world.

Inkling-Small Already Beats Its Bigger Sibling on Some Tests

Thinking Machines is also previewing Inkling-Small, a lighter model with 276 billion total parameters and 12 billion active. On several benchmarks it beats the flagship it’s supposed to sit beneath. Inkling-Small scores 88.3% on GPQA Diamond against 87.2% for Inkling, 46.6% on HLE with tools against Inkling’s 46.0%, and 83.4% on IFBench against 79.8%.

Thinking Machines credits changes to pretraining data and process for the gap. The company plans to publish Inkling-Small’s weights once testing wraps, without a firm date attached.

Third-party access keeps expanding. Databricks already announced day zero access to Inkling’s open weights on its platform, built around its Unity AI Gateway. TogetherAI, Fireworks, Modal and Baseten are expected to follow.

Thinking Machines has not announced pricing for Tinker itself. Developers can already pull Inkling’s weights from Hugging Face or test it inside the Tinker Playground, but what a Tinker subscription eventually costs remains the one number the company still hasn’t shared.

Frequently Asked Questions

What Is Thinking Machines Lab’s Inkling Model?

Inkling is a 975 billion parameter Mixture-of-Experts model that reads text, images and audio using an encoder-free design. Images are processed as 40 by 40 pixel patches through a four-layer hierarchical MLP, and audio is converted into discretized mel spectrograms, both feeding into the same transformer alongside text tokens through a shared, lightweight embedding layer.

Is Inkling Actually Free to Use?

The weights are free to download under Apache 2.0, but a separate terms of use sits alongside that license and narrows what developers can actually do with it, an unusual pairing since Apache 2.0 alone typically carries no such restriction. Running Inkling at any scale still costs money on top of that, since full precision needs roughly 2 TB of GPU memory.

What Is Inkling-Small and When Does It Launch?

Inkling-Small is a preview model that uses the same Mixture-of-Experts design as Inkling, with 276 billion total parameters and 12 billion active. Thinking Machines has not given a release date, saying only that weights arrive once testing concludes.

Can Inkling Run on a Single Consumer GPU?

No. Full precision needs at least eight Nvidia B300 or sixteen H200 accelerators. Thinking Machines built the quantized NVFP4 version specifically for teams without that much GPU capacity, cutting the requirement to roughly four B300s or eight H200s, though that is still well beyond a single consumer card.

How Does Inkling Compare to DeepSeek and GLM?

It depends on the benchmark. Inkling posts 54.3% on SWE-bench Pro Public and leads Kimi K2.6 and DeepSeek V4 Flash Max on the GDPval-AA v2 and Tau-3 banking tests, but GLM 5.2 and DeepSeek V4 Pro score higher on core coding benchmarks like SWE-Bench Verified and Terminal Bench 2.1.

Does Thinking Machines Make Money from Inkling?

Not directly. The company generates revenue through Tinker, its fine-tuning platform. Inkling’s weights themselves are free. Thinking Machines raised a $2 billion seed round at a $12 billion valuation in 2025, before it had released any model or product, and has not disclosed how Inkling’s release affects that funding picture.