Headline: Mira Murati’s Thinking Machines Releases Inkling — a 975B-parameter, Fully Open-Weights Multimodal Model That Could Matter to Crypto Builders
Mira Murati, the former OpenAI CTO, has shipped her first AI model since leaving the company: Inkling. It’s a multimodal, mixture-of-experts model trained from scratch, and its full weights are available openly under an Apache 2.0 license on Hugging Face. For crypto teams and decentralized projects that care about auditability, self-hosting and legal compliance, that last detail is especially notable.
Quick timeline and backstory
- Murati left OpenAI in September 2024 after serving briefly as interim CEO during the Sam Altman board episode in November 2023. She returned to CTO when Altman was reinstated, then departed about ten months later.
- She founded Thinking Machines Lab in February 2025. The startup raised a blockbuster $2 billion seed round at a $12 billion valuation in July 2025, led by Andreessen Horowitz with participation from Nvidia, Accel, ServiceNow, Cisco, AMD and Jane Street. Reports of a $50 billion valuation round surfaced in November 2025 but talks fell apart by January 2026.
- Almost two years after leaving OpenAI, Thinking Machines unveiled Inkling.
What Inkling is — the tech specs
- Architecture: Mixture-of-experts (only part of the network activates per input to keep inference efficient).
- Size: 975 billion total parameters, with 41 billion active per task (not feasible to run on a local laptop).
- Modalities: Text, images, audio (and video in pretraining).
- Context window: 1 million tokens (roughly 750,000 words).
- Training data: Pretrained on 45 trillion tokens across text, images, audio and video.
- Licensing and access: All weights are open and hosted on Hugging Face under Apache 2.0. Murati tweeted: “Our first model, Inkling. Trained from scratch, weights are open, fine-tunable on Tinker today.”
- Fine-tuning: Supported via Thinking Machines’ cloud platform, Tinker.
- Smaller variant: Inkling-Small (276B total / 12B active) — matching the larger model on many reasoning tasks; weights to be released after testing.
Performance and positioning
- Inkling is pitched as a “well-rounded” generalist rather than a narrow specialist.
- Agentic task results:
- MCP Atlas (agent completion rate): Inkling 74.1% — nearly 30 percentage points above Nvidia’s Nemotron 3 Ultra in the same comparison.
- SWE-Bench Verified (autonomously fixing GitHub bugs): Inkling 77.6% vs Nemotron’s 70.7%.
- Safety/robustness:
- FORTRESS Adversarial (handling harmful prompts without over-blocking): Inkling 78.0% — the top score among open-weights models tested.
- Areas where Asian models still lead:
- Terminal Bench 2.1 (autonomous coding agents): Z.ai’s GLM 5.2 scored 82.7% vs Inkling’s 63.8%.
- Kimi K2.6 outperforms on Humanity’s Last Exam (PhD-level scientific reasoning).
- The takeaway: Inkling isn’t the absolute top performer across every benchmark, but it’s the most capable open-weights model released by a Western lab to date. For organizations that can’t or won’t route workloads through models developed in Beijing, Inkling provides a viable, auditable alternative — and fine-tuning can narrow performance gaps for task-specific use cases.
Why this matters for crypto and decentralised projects
- Open weights under Apache 2.0: Enables independent auditing, forensics and on-prem or self-hosted deployments — important for DAOs, regulated crypto firms, and privacy-conscious builders.
- Fine-tunability and Tinker support: Teams can adapt the model for on-chain agent coordination, secure private inference, or custom oracles and bots without relying on closed-source endpoints.
- Large context window and multimodality: Useful for tooling that needs long-form reasoning over contracts, large codebases, multi-modal inputs for NFT metadata, or audio/video data for provenance systems.
Bottom line
Thinking Machines’ Inkling is an important milestone: a Western-built, fully open-weights, large multimodal model with strong agentic and safety results. It won’t dethrone the best Chinese models on every benchmark today, but for crypto developers and enterprises prioritizing auditability, legal clarity and self-hosting, it offers a practical, high-quality alternative — and a platform for building specialized, competitive finetunes.
Read more AI-generated news on: undefined/news
