Who is Mira Murati and why she matters

In November 2023, the tech world was shaken by the news: OpenAI's board of directors ousted CEO Sam Altman. He was temporarily replaced by Mira Murati, the company's then Chief Technology Officer (CTO). Altman returned after five days, but the episode showed just how pivotal a figure Murati was at OpenAI. She was one of the architects behind ChatGPT, GPT-4, and DALL-E. She left OpenAI in September 2024 — with a clear purpose.

Murati soon announced she was founding her own AI company: Thinking Machines Lab. She managed to attract an exceptional team — scientists and engineers who had worked on creating ChatGPT, Character.ai, Mistral, PyTorch, OpenAI Gym, Fairseq, and Segment Anything. Investors weren't far behind: the startup Thinking Machines Lab secured funding of roughly 40 billion Czech crowns (approximately $1.7 billion), despite generating zero revenue at the time the investment round closed. A bet on the team was enough.

Internal turmoil that nearly buried everything

The path to the first model wasn't smooth, however. According to CzechCrunch, Inkling's release was accompanied by severe internal turbulence — a funding crisis, team tensions, and even a concealed workplace romantic relationship that escalated into an attempted internal coup. Thinking Machines Lab hasn't publicly commented on the details, but the very fact that the startup managed to release its first model despite these upheavals testifies to the determination of the founder and the rest of the team to see the project through.

For context: OpenAI experienced similar internal drama in 2023, as did Inflection AI. Startups in the generative AI era operate under immense pressure — time, investor, and human. Thinking Machines Lab weathered the storm, however, and released Inkling at a time when competition from Anthropic, Google DeepMind, Mistral, and Meta Llama is stronger than ever.

Inkling: what exactly Thinking Machines Lab released

Inkling is a multimodal model with a Mixture of Experts (MoE) architecture. This means that instead of one monolithic neural network, the model consists of many specialized "experts," of which only a subset is activated for each query. The result: 975 billion total parameters, but during inference it works with only 41 billion active. This approach reduces computational demands compared to full-size models — a similar logic is used by Mixtral from Mistral AI or GPT-4 (assumed architecture).

What Inkling can do:

• Multimodality: understands text, images, and audio simultaneously — natively, without external add-ons
• 1 million token context window: equivalent to roughly 750,000 words, i.e., several books at once
• Complex reasoning: math, science, coding, tool use (agentic tool use)
• Confidence calibration: the model can estimate how certain it is of its answers — critically important for enterprise deployment
• Adjustable compute budget: can switch between speed and depth of response

A key aspect: Inkling is an open-weights model. The weights are freely available for download via Hugging Face. This puts it in the same category as models like Llama from Meta or Mistral — unlike closed models such as GPT-4o or Claude 3.5 Sonnet, which are only accessible through APIs.

Tinker: fine-tuning as a key strategic pillar

Thinking Machines Lab doesn't just sell a model. Their second product is Tinker — a training API platform designed so that researchers and companies can easily customize Inkling (and other open-source models) for specific use cases.

Tinker uses the LoRA (Low-Rank Adaptation) technique — only a lightweight adapter is trained instead of the model's full weights. The result is nearly comparable to full fine-tuning, but at a fraction of the computational cost. The platform works with four core operations: forward/backward pass, optimization step, token generation, and saving training state.

Pricing is consumption-based (USD per million tokens), with checkpoint storage costing $0.10 per GB per month. Tinker currently supports other popular models as well — Qwen, Kimi-K2.6, DeepSeek-V3.1, and NVIDIA Nemotron. For companies that want to customize AI to their data without building their own infrastructure, it's an attractive option.

Interaction Models: a vision of AI that listens in real time

Alongside Inkling, Thinking Machines Lab published research on so-called Interaction Models — systems capable of genuine two-way communication in real time, across audio, video, and text.

The architecture processes input in 200-millisecond micro-cycles — the model responds continuously, not just after the user finishes their entire sentence. Input (audio as dMel spectrograms, video as 40×40 pixel patches) is processed without prior encoders — everything is trained together. On top of that comes a dual architecture: one model manages real-time interaction, while another runs in the background performing deeper reasoning. Both share context.

This is a direct counterpoint to today's voice assistants that wait for you to finish speaking. If Thinking Machines Lab manages to bring this approach to a product, it could be a fundamental shift in how we communicate with AI.

Philosophy: decentralized AI instead of a single ruler

What markedly sets Thinking Machines Lab apart from OpenAI, Google, or Anthropic is their explicit philosophy. In the blog post "The Future Worth Building Is Human", they reject the vision of centralized, autonomous superintelligence. Instead, they build on a decentralized AI model, where every organization and every person has AI tailored to their own values and needs — not a shared black box controlled by a single corporation.

"AI should empower every organization to excel in its own way, not erase the differences between them," they write on their website. This philosophy directly motivates the open weights of Inkling and the existence of Tinker: it's not just a technical choice, but a political decision about how AI should be distributed in society.

What it means for Czechia and Europe

As an open-weights model, Inkling is accessible to anyone in the world — including Czech developers, researchers, and companies. Direct Czech localization hasn't been confirmed yet, but since it's an open-weights model with fine-tuning support, it's technically entirely feasible to train a Czech language variant — similar to how Czech-tuned versions of Llama or Mistral models are being created.

From the perspective of the EU AI Act, Thinking Machines Lab's open-weights approach is relevant: the European Commission is discussing how to regulate freely available models that cannot be controlled like closed APIs. Inkling will undoubtedly be part of these debates.

For Czech companies and research institutions, Tinker is interesting: if you want to fine-tune a large model on your own company data without having to send data to the OpenAI or Google cloud, it's a legitimate alternative. The cost depends on volume, but the pricing structure is transparent.

Comparison with the competition

Inkling enters a crowded market. How does it stack up at first glance?

• vs. Meta Llama 4: Llama 4 also has an MoE architecture and open weights, but Inkling offers native multimodality including audio
• vs. Mistral: Mistral is strong on efficiency, Inkling goes for a larger parameter scale
• vs. GPT-4o / Claude 3.5: Closed models without the ability to fine-tune your own weights — Inkling offers a key difference here in data control
• vs. DeepSeek V3: The closest competitor in the large MoE open-weights model category — the showdown will be interesting

Thinking Machines Lab hasn't yet published specific benchmark numbers (MMLU, HumanEval, MATH) in the form of a clear comparison table. The company states that Inkling "scores competitively across ten evaluations covering reasoning, agentic tasks, factuality, audio, visual capabilities, and chat." Detailed results are available on the model's page.