Thinking Machines Lab Launches Inkling: Mira Murati’s 975-Billion Parameter Open-Weights AI Model That Challenges Chinese Labs

Key Takeaway: Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, has launched Inkling — a 975-billion-parameter Mixture-of-Experts open-weights model under Apache 2.0. With 41 billion active parameters, a 1-million-token context window, native multimodal support (text, images, audio), and controllable thinking effort, Inkling is the largest American open-weights model to date. It challenges the dominance of Chinese open-source models like DeepSeek V4 and Kimi while giving Western developers a serious alternative to proprietary APIs. Meanwhile, OpenAI revealed GPT-Red, an internal AI super-hacker that autonomously finds vulnerabilities in its own models — succeeding on 84 percent of attack scenarios versus 13 percent for human red-teamers.

July 16, 2026, marks a pivotal moment in the open-source AI landscape. Thinking Machines Lab, the San Francisco-based startup founded by former OpenAI Chief Technology Officer Mira Murati in early 2025, has released its first foundation model — Inkling — under a permissive Apache 2.0 license. At 975 billion total parameters with 41 billion active per token, Inkling is the largest open-weights model to come from a Western lab, directly challenging the Chinese open-source dominance of DeepSeek, Qwen, and Kimi.

The release comes during a remarkable week for AI: OpenAI detailed GPT-Red, an internal AI super-hacker that autonomously finds vulnerabilities in its own models; MIT researchers published GIFT, a framework that teaches vision-language models to generate accurate CAD programs from 2D images; and Cast AI launched Kimchi Coding as a general-availability multi-model coding agent. Together, these developments paint a picture of an AI industry moving faster than ever, with open-weights models closing the gap on proprietary systems.

Inkling: Architecture and Capabilities

Inkling is a Mixture-of-Experts transformer with 975 billion total parameters, of which 41 billion are active for each token generated. The architecture uses 256 routed experts and two shared experts, with six experts active per token. This means Inkling achieves the knowledge capacity of a nearly trillion-parameter model while keeping inference costs comparable to models a fraction of its size.

The model supports a context window of up to 1 million tokens — enough to process entire codebases, lengthy legal documents, or large research papers in a single pass. It was pre-trained from scratch on 45 trillion tokens spanning text, images, audio, and video, using Nvidia GB300 NVL72 clusters. Thinking Machines says the training was inspired by DeepSeek-V3’s MoE architecture but was a from-scratch effort with its own innovations in training methodology and post-training optimization.

A key architectural innovation is controllable thinking effort. Developers can adjust an effort parameter between 0.2 and 0.99, allowing the model to spend more or fewer reasoning tokens depending on the task. According to Thinking Machines, Inkling matches Nvidia’s Nemotron 3 Ultra on Terminal Bench 2.1 while using roughly one-third as many reasoning tokens — a meaningful reduction in inference costs.

Multimodal by Default

Unlike some models that bolt on image or audio understanding after the fact, Inkling reasons natively over text, images, and audio. This means a single model can process a screenshot, answer questions about a podcast, read a document, and write code — all without routing between specialized sub-models. Audio is one of its strongest capabilities: Inkling scores 91.4 percent on VoiceBench and 77.2 percent on MMAU, placing it among the top models for voice-based AI applications.

On coding and agentic tasks, Inkling performs strongly on one-shot web application generation and longer development workflows involving planning, browser interaction, debugging, and iterative improvement. These capabilities reflect the industry-wide shift toward AI agents that can complete real software projects instead of generating isolated code snippets.

Safety and Benchmark Performance

Safety evaluations show Inkling is competitive with the strongest open-weight models: 78.0 percent on FORTRESS Adversarial, 95.9 percent on FORTRESS Benign, and 98.6 percent on StrongREJECT. On reasoning benchmarks, the model scores 78.1 percent on CharXiv RQ (climbing to 82.0 percent when Python tools are enabled).

On the broader benchmark landscape, Thinking Machines positions Inkling as a balanced generalist rather than a model built to dominate every specific evaluation. It consistently ranks alongside today’s strongest open-weight models — between Kimi 2.5 and Kimi 2.6 across reasoning and multimodal tasks — but trails proprietary frontier models from OpenAI and Anthropic on certain specialized benchmarks.

Inkling-Small: Preview of a Lighter Model

Alongside Inkling, Thinking Machines previewed Inkling-Small, a 276-billion-parameter MoE model with 12 billion active parameters. Early results show Inkling-Small performing close to Inkling on reasoning and agentic tasks despite being a fraction of the size. With 12 billion active parameters and controllable thinking effort, Inkling-Small is designed for high-volume, cost-sensitive workloads such as code review, LLM-as-judge grading, and synthetic data generation. Full weights will be released once testing is complete.

Availability and Deployment

Inkling’s full weights are available on Hugging Face as both the original checkpoint and an NVFP4 quantized checkpoint optimized for Nvidia Blackwell GPUs. The quantized version requires approximately half the GPU memory, making it feasible on eight Nvidia H200 GPUs instead of the 16-plus required for full precision.

For developers who prefer API access, Inkling is available through Thinking Machines’ own Tinker platform (which offers fine-tuning capabilities) and through third-party providers including TogetherAI, Fireworks, Modal, Databricks, and Baseten. The model supports SGLang, vLLM, Llama.cpp, TokenSpeed, and Hugging Face Transformers — broad inference engine support that makes integration straightforward.

Pricing is competitive: Thinking Machines is offering a limited-time 50 percent discount on Inkling API access, with full pricing available in documentation. The company also provides an Inkling Playground in the Tinker console for free evaluation.

OpenAI GPT-Red: The AI That Hunts AI Vulnerabilities

On July 15, OpenAI published details of GPT-Red, an internal AI system built to autonomously attack its own models and surface prompt injection vulnerabilities. GPT-Red was trained using self-play reinforcement learning, where it acts as the attacker against defender models across varied scenarios. As defenses improve, GPT-Red is forced to invent harder attacks in a continuous adversarial loop.

The results are striking. GPT-Red succeeds on 84 percent of attack scenarios against 13 percent for human red-teamers. It cuts direct prompt injection failures to one-sixth the rate of OpenAI’s best production model from four months earlier. A class of “fake chain-of-thought” attacks that worked more than 95 percent of the time against GPT-5.1 now succeeds less than 10 percent of the time against GPT-5.6 — thanks to GPT-Red’s adversarial training.

OpenAI is keeping GPT-Red internal. The company said it will not release the model, as the attack capabilities it develops could be dangerous in public hands. Precursor versions have been used in training since GPT-5.3.

“GPT-Red is not a product and will not be released,” OpenAI stated. The system has limits — it struggles with multi-turn conversational attacks and image-based prompt injection — but it represents a significant step toward automated AI safety testing.

MIT GIFT: Teaching AI to Generate CAD From 2D Images

In a development particularly relevant for engineering and manufacturing readers, MIT researchers presented a system called GIFT (Geometric Inference Feedback Tuning) at the International Conference on Machine Learning. GIFT teaches vision-language models to convert 2D design images into accurate CAD programs with only 20 percent of the computation required by competing approaches.

The system works by testing the model on CAD generation problems, identifying near-miss solutions, and correcting them into successful ones. These corrected solutions become training data that teaches the model to overcome its weaknesses. For rapid prototyping and design iteration, this could substantially reduce the time from concept to 3D-printable model — a capability every engineer will want to watch closely.

Cast AI Kimchi Coding: Multi-Model Coding Agent Hits GA

Cast AI announced general availability of Kimchi Coding on July 15, an autonomous multi-model coding agent that routes each task to the best-fit model based on complexity and cost. In shadow-mode evaluations, Kimchi matches or exceeds the quality of single-commercial-model coding agents while costing 2.5 times less, achieved by routing most work to open-weight models and running self-hosted inference on optimized GPU infrastructure.

What makes Kimchi relevant alongside Inkling is the same trend: the AI industry is moving toward multi-model orchestration, where the question is no longer “which model is best” but “how do I route each subtask to the model that balances quality and cost optimally.” Inkling, as a strong general-purpose open model, becomes a natural component in such systems.

What This Means for AI Development

The simultaneous release of Inkling, GPT-Red details, MIT GIFT, and Kimchi Coding in a single week signals an accelerating AI landscape where the boundaries between proprietary and open-source continue to blur. Inkling gives Western developers a credible alternative to Chinese open-source models and proprietary APIs — one that can be self-hosted, fine-tuned, and customized without licensing restrictions under Apache 2.0.

For enterprises evaluating AI infrastructure, the implications are clear: the open-source AI ecosystem is no longer playing catch-up. With models like Inkling offering 1-million-token context windows, native multimodality, and competitive benchmark scores at zero licensing cost, the total cost of ownership calculation for AI systems increasingly favors open-weights solutions — especially for high-volume, latency-sensitive, or data-sovereignty-constrained deployments.

Frequently Asked Questions

What is Thinking Machines Lab’s Inkling model?

Inkling is a 975-billion-parameter Mixture-of-Experts open-weights AI model released under Apache 2.0. It has 41 billion active parameters per token, a 1-million-token context window, and native support for text, images, and audio. It is the first foundation model from Mira Murati’s AI startup.

How does Inkling compare to GPT-5.6 and Claude?

Inkling trails proprietary frontier models on certain specialized benchmarks but is competitive with the strongest open-weight models. Its key advantages are Apache 2.0 licensing (commercial use without restrictions), controllable thinking effort for cost optimization, and native multimodality.

What hardware is needed to run Inkling?

Full 16-bit precision requires approximately 2TB of GPU memory (16 Nvidia H200s or 8 B300s). An NVFP4 quantized version is available that requires roughly half the hardware. For lighter workloads, Inkling-Small (12B active parameters) is in preview.

What is OpenAI’s GPT-Red?

GPT-Red is an internal OpenAI system that autonomously attacks its own models to find security vulnerabilities. It succeeds on 84 percent of attack scenarios compared to 13 percent for human red-teamers. It will not be released publicly.

Can I use Inkling commercially?

Yes. Inkling is released under Apache 2.0, which permits commercial use, modification, and redistribution without licensing fees.

Where can I access Inkling?

The model weights are available on Hugging Face. API access is available through Thinking Machines’ Tinker platform, TogetherAI, Fireworks, Modal, Databricks, and Baseten.

Related Reading

• GPT-5.6 Family Explained: Sol, Terra and Luna
• ChatGPT Work: OpenAI’s Autonomous Agent
• Open-Source AI Models: Tencent Hy3, DeepSeek-V4-Flash Compared

Sources

• Thinking Machines Lab — Introducing Inkling
• The Register — Thinking Machines’ Inkling Analysis
• MIT Technology Review — GPT-Red: OpenAI’s LLM Super-Hacker
• MIT News — GIFT: 2D to 3D CAD Generation
• Cast AI — Kimchi Coding General Availability
• Tech Startups — Inkling vs Kimi and Nemotron

Disclosure: Some links in this article are affiliate links. If you purchase through these links, we may earn a small commission at no additional cost to you.