Thinking Machines Lab's Inkling: A 975B-Parameter Open-Weights Multimodal MoE with Controllable Thinking Effort

Thinking Machines Lab's Inkling, released on July 15, 2026, marks a significant entry into the highly competitive frontier of open-weights multimodal artificial intelligence models, boasting a 975 billion-parameter Mixture-of-Experts (MoE) architecture with 41 billion active parameters and an expansive 1 million-token context window. This debut model, trained from scratch and released under an Apache 2.0 license, uniquely incorporates "controllable thinking effort," allowing developers to balance inference speed with reasoning depth. Inkling accepts text, image, and audio inputs, generating UTF-8 text outputs, and is designed for diverse applications including agentic systems, coding assistants, chatbots, and retrieval-augmented generation (RAG).

The significance of Inkling's release lies in its combination of cutting-edge architectural choices and its open-weights distribution, poised to disrupt the enterprise AI landscape. The MoE architecture, where each token is routed to 6 of 256 specialized experts plus 2 shared experts, enables efficient inference by activating only a fraction of the total parameters, addressing the computational demands typically associated with such massive models. This efficiency is crucial for developers seeking to deploy powerful AI on their own infrastructure, as the full BF16 checkpoint alone requires at least 2 TB of aggregated VRAM, though quantized versions can run on significantly less. The Apache 2.0 license further democratizes access, providing developers and enterprises with the flexibility to fine-tune Inkling on proprietary data, integrate it into third-party products, and deploy it without per-token API fees, fostering innovation and mitigating vendor lock-in. This open approach facilitates customization, control, and compliance, which are paramount for businesses in regulated industries.

Inkling's "controllable thinking effort" is a notable differentiator, allowing users to adjust the model's reasoning intensity from a swept range of 0.2 to 0.99, or through named levels like "none," "minimal," "low," "medium," "high," "xhigh," and "max". This feature directly addresses a critical challenge in AI deployment: optimizing cost and latency for varied tasks. Simple queries can be processed quickly with minimal effort, while complex problems requiring multi-step planning, coding, or mathematical reasoning can leverage higher effort levels for improved accuracy, albeit at increased token usage and latency. This granular control aligns with the industry's shift towards routing tasks to models or effort levels appropriate for their difficulty, moving away from the "bigger is better" mentality that often led to overpaying for unnecessary computation. Inkling has demonstrated competitive performance in certain benchmarks, achieving 77.6% on SWE-bench Verified, surpassing Nvidia Nemotron 3's 71.9%, and scoring 91.4% on VoiceBench. Its FORTRESS Adversarial score of 78.0% also leads among compared open-weights models.

In the current AI landscape of mid-2026, Inkling enters a fiercely competitive arena populated by both proprietary and open-weight models. Proprietary leaders like Anthropic's Claude Fable 5, Claude Mythos 5, Google's Gemini 3.1 Pro, and OpenAI's GPT-5.4 continue to lead overall benchmarks, particularly in complex reasoning and unified multimodal understanding. Claude Fable 5, for instance, leads proprietary rankings with a score of 91, excelling in agentic and coding tasks. However, open-weight models have significantly closed the capability gap, especially in coding and reasoning. Rivals such as Zhipu AI's GLM-5.2 offer a 1 million-token context window and multiple reasoning-effort levels, positioning it as a strong general-purpose open-weight model for long-horizon agentic work. DeepSeek V4 Pro, with 1.6 trillion total parameters and 49 billion active, also provides a 1 million-token context, excelling in knowledge, mathematics, and reasoning. While multimodal capabilities still see a lead from closed-source models like GPT-5.4 and Gemini 3.1 Pro for unified text-image-audio-video workloads, open-weight models like MiMo V2 Omni and Qwen 3.5-Omni are competitive in image and audio processing. Inkling's native multimodal processing of text, image, and audio into a shared hidden space, processed jointly by a 66-layer decoder-only transformer, places it firmly in this advanced category.

Looking ahead, Inkling's open-weights model, combined with its "controllable thinking effort" and multimodal capabilities, positions it as a strong enabler for the next wave of AI applications. The ability for enterprises to fine-tune and run this powerful model on their own infrastructure, with day-zero support across various inference engines and platforms like Databricks, will accelerate the development of specialized, privacy-preserving AI agents and applications. This caters to a growing demand for domain-specific models and agentic workflows that require deep integration with proprietary data and robust governance. The release of a smaller preview model, Inkling-Small, further suggests a strategy to cater to a broader range of deployment scenarios, including those with more constrained hardware. The future will likely see increased adoption of hybrid AI strategies, where enterprises strategically combine open-weight models like Inkling for customized, cost-effective, and auditable operations with proprietary models for tasks demanding state-of-the-art general intelligence or highly specialized capabilities not yet matched by open alternatives. Thinking Machines Lab's focus on "resistance to censorship" also highlights a growing demand for AI models that offer unbiased factual outputs, a crucial consideration for enterprise trust and integrity.