Thinking Machines Lab's Inkling: An Open-Weight AI Challenger

TL;DR

Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, has released its first model, Inkling. This open-weight, 975-billion-parameter mixture-of-experts model is designed to be a strong, customizable foundation for enterprises. It supports text, images, audio, and video natively, and introduces controllable thinking effort to balance performance with cost. While not the absolute leader on every benchmark, Inkling's strategic positioning—as a base for fine-tuning on the company's Tinker platform—represents a direct bet against the one-size-fits-all approach of closed models from OpenAI, Anthropic, and Google.

Inkling: A New Open-Weight Contender

Thinking Machines Lab has officially entered the AI arena with Inkling, an open-weight model that prioritizes flexibility and real-world utility over raw benchmark dominance. The company, which emerged from stealth just nine months ago, is making a clear statement: the future of AI is not about a single, monolithic model, but about models that can be adapted and owned by the organizations that use them.

Inkling is a Mixture-of-Experts (MoE) transformer with 975 billion total parameters, but it activates only 41 billion for any given task. This design, which mirrors the architecture of China's DeepSeek-V3, allows for high capability without prohibitive computational cost. The model was pretrained from scratch on 45 trillion tokens spanning text, images, audio, and video, giving it native multimodal reasoning abilities.

The Strategic Bet on Customization

Murati's team is not trying to beat GPT-5.6 or Claude Opus 4.8 on every leaderboard. Instead, Inkling is positioned as a base model for fine-tuning. The company's Tinker platform is the centerpiece of this strategy, offering developers a playground to chat with Inkling and a suite of tools to customize it for specialized tasks. This is a direct response to a growing enterprise demand for models that can be tailored with proprietary data without the risk of leaking sensitive information to a closed API provider.

As Microsoft CEO Satya Nadella recently noted, enterprises using closed AI models are effectively paying twice: once for the subscription, and again by handing over their confidential data to improve the model. Inkling's open-weight nature allows organizations to maintain full control, making it a compelling option for sectors like finance, healthcare, and defense where data privacy is paramount.

Technical Deep Dive: Architecture and Training

Inkling's architecture is a deliberate evolution of proven MoE techniques. It uses 256 routed experts and 2 shared experts per MoE layer, with a sigmoid-based router for load balancing. A notable departure from common practice is its use of relative positional embeddings instead of Rotary Position Embeddings (RoPE), which the team found performs better at longer sequence lengths. The model supports a context window of up to 1 million tokens.

Training was a massive undertaking. Pre-training on 45 trillion tokens was followed by a post-training phase that included supervised fine-tuning (SFT) on synthetic data generated by other open-weight models, including China's Kimi K2.5. The bulk of the compute, however, went into large-scale reinforcement learning (RL). The team scaled RL to over 30 million rollouts, observing log-linear improvements in reasoning performance throughout the process. This RL training also led to an emergent compression of the model's chain of thought, making it more efficient without sacrificing accuracy.

Performance and Positioning

On benchmarks, Inkling holds its own. It scores 77.6% on SWE-bench Verified, 97.1% on AIME 2026, and 87.2% on GPQA Diamond. In safety, it achieves 78% on the adversarial FORTRESS benchmark, the highest among open-weight models tested. Its audio and vision capabilities are also strong, placing it among the best open-weight models for multimodal tasks.

However, the company is transparent about its limitations. Inkling is not the strongest model in any single category. On HLE (text only), it scores 29.7%, well behind Claude Fable 5's 53.3%. The strategy is not to win every benchmark, but to offer a balanced, efficient, and customizable foundation. The introduction of controllable thinking effort is key here: developers can dial the model's reasoning depth up or down, trading accuracy for speed and cost as needed.

The Geopolitical Angle

Inkling's release also carries geopolitical weight. The best open-weight models today largely come from China, and Thinking Machines' decision to use Chinese models like Kimi K2.5 for synthetic data has drawn scrutiny. Critics point to a double standard: Chinese use of US models is often labeled theft, while US adoption of Chinese open models is framed as engineering. Murati's team is effectively filling a vacuum created by US protectionist policies, offering a domestic alternative that is competitive with, and in some respects superior to, the Chinese models that American companies are increasingly pressured to avoid.

Availability and Ecosystem

Inkling is available now on the Tinker platform with context length options of 64K and 256K tokens. The company is offering a 50% discount for a limited time. Full weights are available on Hugging Face, and the model is supported on inference platforms including Together AI, Fireworks, Modal, Databricks, and Baseten. The company has also partnered with RadixArk for SGLang support, Inferact for vLLM, and Unsloth for llama.cpp.

Alongside Inkling, the company previewed Inkling-Small, a 276B-parameter model (12B active) that matches its larger sibling on many reasoning and agentic tasks, offering a compelling option for latency-sensitive workloads.

Why It Matters

Thinking Machines Lab's Inkling is not just another model release; it is a strategic manifesto. It argues that the future of enterprise AI lies not in renting intelligence from a few gatekeepers, but in owning and customizing it. By releasing a competitive open-weight model and building a platform around fine-tuning, Murati is betting that the market will value control and adaptability over raw, untailored power. If she is right, Inkling could be the first major step in a new paradigm for how businesses deploy AI.