Mira Murati—the high-profile former CTO and interim CEO of OpenAI—has shipped her first major product since leaving the company: a fully open-source, large multimodal model called Inkling from her startup, Thinking Machines Lab.
What was released
- Inkling is a mixture-of-experts model trained from scratch and released with all weights available for free on Hugging Face under an Apache 2.0 license. Murati announced it on X: “Our first model, Inkling. Trained from scratch, weights are open, fine-tunable on Tinker today.”
- The model is huge: 975 billion total parameters, with about 41 billion active per task (so it’s meant for cloud/inference servers, not local laptops).
- Multimodal inputs: text, images, and audio. Context window: 1 million tokens (roughly 750,000 words).
- Pretraining corpus: roughly 45 trillion tokens across text, image, audio, and video.
- Fine-tuning and managed workflows are supported on Thinking Machines’ Tinker cloud platform; full weights are downloadable with no restrictions.
Background and company trajectory
- Murati left OpenAI in September 2024 after serving as CTO and briefly as interim CEO during the November 2023 board crisis around Sam Altman.
- She founded Thinking Machines Lab in February 2025. The startup raised a massive $2 billion seed in July 2025 at a $12 billion valuation—led by Andreessen Horowitz and joined by Nvidia, Accel, ServiceNow, Cisco, AMD, and Jane Street—one of the largest seed raises in Silicon Valley history.
- Reports that Thinking Machines pursued a follow-on at a $50 billion valuation surfaced in November 2025, but those talks reportedly collapsed by January 2026.
How Inkling performs (benchmarks)
- Inkling shines on agentic benchmarks—tasks that require connecting to external tools and autonomously completing real-world workflows.
- MCP Atlas (agentic task completion): Inkling scores 74.1%, described as nearly 30 points above Nvidia’s Nemotron 3 Ultra in the same comparison.
- SWE-Bench Verified (autonomous fixing of GitHub issues): Inkling 77.6% vs Nemotron’s 70.7%.
- FORTRESS Adversarial (handling harmful prompts without over-blocking): Inkling 78.0%, the highest among open-weights models tested.
- Areas where Inkling trails top Chinese models:
- Terminal Bench 2.1 (autonomous coding in real terminal): Z.ai’s GLM 5.2 scores 82.7% vs Inkling’s 63.8%.
- Humanity’s Last Exam (PhD-level scientific reasoning): Kimi K2.6 leads.
- Bottom line: Inkling is not the single strongest model overall, but it’s the most capable open-weights model produced by a Western lab so far—particularly strong for agentic, tool-using workflows.
Why this matters for crypto, Web3, and privacy-conscious developers
- Open weights under Apache 2.0 are a big deal for builders who need to self-host, audit, or fork models for legal, compliance, or privacy reasons—common requirements in crypto and enterprise blockchain projects.
- The mixture-of-experts design and multimodal capabilities make Inkling a candidate for decentralized or hybrid architectures: on-chain or off-chain oracles, private inference, local fine-tuning for compliance, and customized agents that interact with smart contracts and external services.
- Tinker’s fine-tuning platform plus downloadable weights means teams can iterate models on private data or adapt Inkling to specialized dApps and tooling without routing sensitive data through a closed provider.
A smaller sibling and release cadence
- Thinking Machines also previewed Inkling-Small: 276 billion total parameters with 12 billion active, which reportedly matches the larger model on many reasoning benchmarks. Its weights will be released after testing, with no specific timeline given.
Takeaway
Inkling represents a clear moment for the open AI ecosystem: a Western lab with deep capital backing has shipped a large, multimodal, open-weights model aimed at agentic tasks and enterprise use cases. For crypto builders and privacy-first teams, that means a viable, ready-to-adapt alternative to Chinese-open models—or closed-source Western offerings—backed by a startup with significant funding and ambition.
Read more AI-generated news on: undefined/news
