Inkling: Thinking Machines Lab Unleashes Open-Weights Multimodal AI with Controllable Effort

Introducing Inkling: A New Era of Customizable Multimodal AI

At a glance, The artificial intelligence landscape is constantly evolving, with new models pushing the boundaries of what’s possible. Thinking Machines Lab has made a significant splash with the release of Inkling, their first model trained entirely from scratch. This isn’t just another large language model; Inkling is a powerful, open-weights multimodal Mixture-of-Experts (MoE) transformer designed for unparalleled customization and control.

Meanwhile, Pitched as a foundational base for bespoke AI solutions, Inkling offers a unique combination of scale, multimodality, and a novel feature: controllable thinking effort. This allows developers and researchers to fine-tune the model’s computational expenditure, making it a highly efficient and adaptable tool for a diverse range of applications.

What Makes Inkling Stand Out?

At its core, Inkling is a sophisticated MoE transformer boasting an impressive 975 billion total parameters, with 41 billion active parameters during inference. Here are some of its key specifications:

Architecture: Mixture-of-Experts (MoE) transformer.
Parameter Count: 975 billion total, 41 billion active.
Context Window: Supports up to 1 million tokens, enabling deep and extensive contextual understanding.
Multimodality: Accepts text, images, and audio as input, providing a comprehensive understanding of various data types.
Output: Generates high-quality UTF-8 text.
Training Data: Pretrained on an enormous dataset of 45 trillion tokens, encompassing text, images, audio, and video.
Open Weights: Released with open weights, making it fine-tunable and highly customizable on platforms like Tinker.

In practical terms, Thinking Machines Lab also teased Inkling-Small, a more compact MoE with 276 billion total parameters and 12 billion active. This smaller sibling has shown performance comparable to, or even exceeding, its larger counterpart on many benchmarks, with its weights slated for release post-testing.

Delving into Inkling’s Advanced Architecture

The power of Inkling lies in its meticulously crafted architecture. It features a 66-layer decoder-only transformer, underpinned by a sparse MoE feed-forward backbone. Each MoE layer is equipped with 256 routed experts and 2 shared experts.

For every token, six routed experts and both shared experts are activated, ensuring a dynamic and efficient processing pipeline. The expert selection is managed by a sigmoid-based router, which incorporates an auxiliary-loss-free load-balancing bias for optimal distribution.

Innovative Attention Mechanism

For example, Inkling’s attention mechanism departs from conventional designs. It employs an interleaved structure of sliding-window and global layers at a 5:1 ratio, utilizing 8 Key-Value (KV) heads.

Instead of standard RoPE, the model leverages a relative positional embedding, which the lab reports offers superior extrapolation capabilities. Additionally, short convolutions are applied after key and value projections, as well as on residual branch outputs, further enhancing its processing capabilities.

Encoder-Free Multimodality

One of Inkling’s most intriguing architectural choices is its encoder-free approach to multimodality. Audio inputs are transformed into dMel spectrograms, while images are processed into 40×40 pixel patches via a four-layer hMLP. A lightweight embedding layer then projects both these modalities, allowing the decoder to process them seamlessly alongside text tokens.

Controllable Thinking Effort: A Game Changer

That said, Perhaps Inkling’s most innovative feature is its controllable thinking effort. During its asynchronous Reinforcement Learning (RL) training, the research team manipulated the system message and per-token cost, enabling the model to learn how to allocate different token budgets for various tasks. This translates into a direct control surface for users, exposed as a reasoning_effort argument with named levels (e.g., ‘minimal’, ‘medium’, ‘max’).

This feature means that cost and latency can be tuned per API call, rather than being fixed per model. Inkling can achieve the same Terminal Bench 2.1 performance as Nemotron 3 Ultra while using only a third of the tokens, showcasing remarkable efficiency.

Interestingly, Alongside this efficiency, the team also prioritized trustworthiness, aiming for a model that is both powerful and reliable.

Performance Benchmarks: A Competitive Edge

Evaluated at an effort of 0.99 and temperature 1.0, Inkling demonstrates highly competitive performance against its open-weights peers. Notably, it leads the pack on the challenging FORTRESS Adversarial benchmark with a score of 78.0%. While it excels in some areas, it also shows room for improvement in others, trailing models like GLM 5.2 on Terminal Bench 2.1.

Key performance highlights include:

FORTRESS Adversarial: 78.0% (leading among compared open-weights models).
MMMU Pro: 73.5%.
VoiceBench: 91.4%.
Design Arena’s Agentic Web Dev leaderboard: 1257 (blinded human evaluation).

Notably while Inkling’s efficiency is a major strength, its BF16 configuration demands significant VRAM (2 TB aggregated), though an NVFP4 version significantly reduces this to 600 GB.

Deployment and Fine-Tuning Made Accessible

Meanwhile, Thinking Machines Lab has made Inkling highly accessible for deployment and fine-tuning. Two main checkpoints are available:

BF16: Requires substantial VRAM, typically 8x NVIDIA B300 or 16x H200.
NVFP4: More efficient, running on 4x B300 or 8x H200 with W4A4 or W4A16 configurations.

Inkling integrates seamlessly with popular runtimes such as SGLang, vLLM, TokenSpeed, Unsloth, and Hugging Face transformers. For fine-tuning, the model is available on Tinker with various context options, and tml-renderers are provided for post-training with multimodal inputs and tool calls. Furthermore, hosted APIs are available through major platforms like TogetherAI, Fireworks, Modal, Databricks, and Baseten, simplifying its integration into existing workflows.

Practical Applications and Use Cases

Inkling’s unique capabilities open doors for several compelling use cases:

Voice-and-Vision Agents

A primary design goal was to support sophisticated interaction models. Imagine a customer support agent powered by Inkling, capable of ingesting a 16kHz WAV audio call alongside a screenshot, and then autonomously generating a structured support ticket.
Cost-Tiered Agentic Pipelines

For example, The controllable effort feature allows for dynamic resource allocation. Low-effort settings can handle initial routing and triage, while max-effort can be reserved for complex problem-solving or repair steps within the same deployment, optimizing compute costs.
Domain Fine-Tuning

The lab highlights its success in financial judgment tasks, where fine-tuning Inkling helped bridge the gap from a generalist model to a domain-specific expert. Its strong performance on chart-heavy analytics (82.0% on CharXiv RQ with Python) also makes it ideal for data-intensive fields.

Key Strengths and Considerations

Strengths

Open & Flexible: Apache 2.0 weights, 1M-token context, native text, image, and audio input.
Efficient Performance: Controllable effort allows matching Nemotron 3 Ultra’s performance at a third of the token cost.
Robust Against Adversaries: Achieves the highest FORTRESS Adversarial score (78.0%) among compared open-weights models.
Broad Support: Day-0 support across major transformer frameworks and five hosted APIs.
Enhanced Decoding: Ships with a multi-token-prediction drafter for speculative decoding.

Considerations

Resource Intensive: BF16 deployment requires significant aggregated VRAM (2 TB).
Benchmark Gaps: Trails some competitors like GLM 5.2 and Kimi K2.6 on benchmarks such as HLE and SWEBench Verified.
Limited Output: Currently, Inkling only produces text output, with no audio or image generation.
Unreleased Small Model: Inkling-Small weights are not yet publicly available.
Safety Risks: The project page acknowledges residual safety risks related to role-play and indirect prompts.

The Future of Customizable AI

That said, Inkling represents a significant step forward in the realm of open-weights multimodal AI. By offering deep customization, flexible deployment options, and the innovative controllable thinking effort, Thinking Machines Lab has equipped developers with a powerful new tool. As the AI community continues to push boundaries, Inkling’s focus on efficiency and adaptability positions it as a key player in shaping the next generation of intelligent applications.

Expert Perspective

A practical read on Inkling AI Model starts with inkling. That is where the earliest effects are likely to show up if this development keeps building.

What happens next will come down to adoption speed, policy response, and execution quality. That combination could make Inkling AI Model a meaningful reference point across model.

For decision-makers, the useful lens is not the headline alone but how weights changes priorities once organizations have to respond.

Frequently Asked Questions

Why is Inkling AI Model important?

Introducing Inkling: A New Era of Customizable Multimodal AIAt a glance, The artificial intelligence landscape is constantly evolving, with new models pushing the boundaries of what’s possible.

What impact could Inkling AI Model have?

Thinking Machines Lab has made a significant splash with the release of Inkling, their first model trained entirely from scratch.

What should readers watch next with Inkling AI Model?

This isn’t just another large language model; Inkling is a powerful, open-weights multimodal Mixture-of-Experts (MoE) transformer designed for unparalleled customization and control.Meanwhile, Pitched as a foundational base for bespoke AI solutions, Inkling offers a unique combination of scale, multimodality, and a novel feature: controllable thinking effort.