Inkling: Our Open-Weights Model
Thinking Machines Lab has launched Inkling, an open-weights, multimodal Mixture-of-Experts AI model, touting a 1M token context and controllable thinking effort. This release sparks avid discussion on its competitive standing against other frontier open models and the elusive business models underpinning open-weights AI. Hacker News is abuzz with comparisons to GLM and DeepSeek, alongside the perennial question of how to fund such ambitious, publicly available research.
The Lowdown
Thinking Machines Lab has introduced Inkling, its first open-weights, multimodal Mixture-of-Experts (MoE) AI model, advancing its mission to build AI that extends human capabilities. Designed as a flexible foundation for customization, Inkling aims to balance performance with cost-efficiency.
- Inkling is an MoE transformer with 975 billion total parameters (41B active) and supports a context window of up to 1 million tokens.
- It's multimodal, natively reasoning over text, images, and audio, and pretrained on 45 trillion tokens.
- Key capabilities include agentic coding, tool use, controllable thinking effort (balancing performance with token efficiency), and strong "epistemics" (calibration, instruction following, censorship resistance) and safety.
- The architecture features a sigmoid-based router for its 256 routed and 2 shared experts, interleaved sliding-window and global attention, and a relative positional embedding.
- Training involved a hybrid optimization strategy, large-scale asynchronous Reinforcement Learning (RL) over 30 million rollouts, which led to a log-linear improvement in reasoning and an emergent compression of the chain of thought.
- A smaller preview model, Inkling-Small (12B active parameters), is also announced, matching or exceeding Inkling on many benchmarks due to improved training.
- Inkling is available for fine-tuning on Thinking Machines' Tinker platform, offering a playground, cookbook recipes, and integrations with numerous inference providers (Together, Fireworks, Modal, Databricks, Baseten, etc.) and open-source inference libraries (SGLang, Miles, vLLM, TokenSpeed, llama.cpp). Its full weights are available on Hugging Face.
Inkling is positioned as a broad, balanced generalist model, emphasizing its customizability and practical application in real-world scenarios, with future iterations expected to enhance its multimodal capabilities further.
The Gossip
Benchmark Brawls & Practical Performance
Commenters are quick to benchmark Inkling against established open-weights models like GLM 5.2, Nemotron, and DeepSeek, often noting it performs well but isn't always superior, especially in agentic coding, despite its size. Skepticism arises regarding radar chart visualizations and whether benchmarks truly reflect real-world utility. Many praise its multimodal capabilities and long context, anticipating the smaller Inkling-Small as a potentially game-changing, cost-effective option.
Open-Weights' Operating Quandary
A prevailing question among the community is the viability of a business model for companies releasing open-weights AI. Discussions center on how Thinking Machines Lab plans to monetize Inkling, with fine-tuning services (Tinker) and API hosting being proposed revenue streams. There's speculation that these efforts are "VC charity" or a strategy for mindshare and market presence rather than immediate profit, especially given intense competition and the absence of a clear moat.
American AI Ambitions
Several users voice a strong desire for robust, American-made open-source AI models, seeing Inkling and Thinking Machines Lab as a potential answer to the dominance of models from other countries, particularly China (e.g., DeepSeek, Z.ai). This reflects a sentiment of national competition and a hope for increased diversity and innovation within the open-source AI landscape from US-based entities.