Thinking Machines Lab Releases Inkling: A 975B-Parameter Open-Weights Multimodal MoE With 41B Active Parameters And Controllable Thinking Effort
Evolving story · 2 updatesThinking Machines' Open-Weight Model, InklingTimeline →Thinking Machines Lab launched Inkling, a 975B-parameter open-weights multimodal model featuring controllable thinking effort and Apache 2.0 licensing.
- Inkling is a 975B-parameter open-weights model released under Apache 2.0.
- The Mixture-of-Experts architecture uses 41B active parameters per token.
- Developers can control the thinking effort to balance cost and reasoning depth.
- The model supports text, image, and audio inputs with a 1M token context.
Thinking Machines Lab has introduced Inkling, its first foundation model trained entirely from scratch. The release features a massive architecture with 975 billion total parameters, utilizing a Mixture-of-Experts design to keep active parameters at 41 billion during inference.
The model is licensed under Apache 2.0, allowing for broad commercial use and modification. It supports a 1 million token context window and processes text, images, and audio natively. The lab explicitly states that Inkling is not designed to be the most powerful model available.
Instead, the focus is on customization and control. The standout feature is the ability to adjust the thinking effort of the model, giving developers granular control over inference cost and reasoning depth for specific applications.
Provides access to a massive, permissively licensed model with unique inference controls.
Apache 2.0 licensing allows for unrestricted commercial integration and customization.
Signals a competitive shift towards controllable infrastructure over raw benchmark performance.
- Mixture-of-Experts (MoE)
- A neural network architecture that activates only a subset of parameters for each input, increasing efficiency.
Open SourceLinus Torvalds Reaffirms That Linux Is Not "Anti-AI" And Not A "Social Warrior" Project
Open SourceI Built a Task Orchestrator, Then Deleted Its Best Number
Open SourceGerman AI consortium releases Soofi S, an open 30B model that tops benchmarks in both English and German
Open Source🇮🇳 ForThePeople.in: One Engineer’s Quest to Put Every Indian District’s Data in Your Pocket
Chinese AI startup MiniMax plans to open-source a 2.7 trillion parameter model later this year
AI ToolsLLM Evals For Developer Tools: Useful, Correct, Safe
A new guide helps developers evaluate the usefulness, correctness, and safety of LLM features in their tools. The guide covers LLM-powered code suggestions and code fixes.
BusinessMicrosoft is reportedly training salespeople to talk down OpenAI and Anthropic
Microsoft is reportedly instructing its sales teams to emphasize the cost-efficiency and performance of its own internal AI models over competitors like OpenAI and Anthropic.
xai-org/grok-build, now open source
xAI released the source code for its Grok CLI tool after users discovered it uploaded entire directories, including sensitive files like SSH keys, to its servers.
AI ToolsSmash Story: The Demo Script That Out-Debugged My Test Suite
A live demo script caught an API-contract bug that was missed by automated tests. The bug was only revealed when the demo was run in a production environment.
Nvidia to partner with Toyota's Woven City on physical AI tech - Nikkei Asia
Nvidia is partnering with Toyota's Woven City to develop physical AI technologies for robotics and smart city infrastructure.
Japan’s Enterprises and Startups Build Industry-Specialized AI With NVIDIA Nemotron Open Models - NVIDIA Newsroom
NVIDIA has announced that Japanese enterprises and startups are leveraging its Nemotron open models to build industry-specialized AI solutions.