Thinking Machines Lab Releases Inkling Open-Weight Model
Thinking Machines Lab released Inkling on July 15, 2026, its first general-purpose model family with downloadable weights under Apache 2.0. The company says the main mixture-of-experts model has 975 billion total parameters, 41 billion active per token, multimodal inputs, and a one-million-token context window; independent performance testing is still needed.
Thinking Machines Lab released Inkling on July 15, 2026, its first general-purpose model family with downloadable weights. The company's launch announcement says Inkling was trained from scratch, while its model card lists an Apache 2.0 license and support for text, image, and audio inputs with text output.
The main model is a mixture-of-experts transformer with 975 billion total parameters and 41 billion active per token. Thinking Machines Lab says it was pretrained on 45 trillion tokens spanning text, images, audio, and video and supports a context window of up to one million tokens. The company also released a preview of Inkling-Small with 12 billion active parameters.
A model built for customization
Thinking Machines Lab describes Inkling as a broad model trained across reasoning, coding, tool use, instruction following, factuality, vision, and audio tasks. The model card identifies a 66-layer decoder-only architecture in which each token is routed through a subset of experts, and presents the release as suitable for fine-tuning and integration into third-party applications.
The company's benchmark charts are vendor-reported results, not an independent evaluation. Reuters reported that competing open and closed models remained ahead overall, while Inkling showed competitive results on some agent-related tasks. Those comparisons establish an evaluation starting point, not a guarantee of performance on a specific workload.
What practitioners should test
Open weights give engineering teams more control over deployment and adaptation than an API-only model, but the headline specifications do not determine production cost by themselves. Active parameter count affects compute per token, while the full weight footprint, serving engine, quantization, batching, and context length all affect memory requirements, throughput, and latency.
The one-million-token figure is a supported maximum context length, not evidence that every long-context task will remain accurate or economical at that limit. Teams considering Inkling should test task quality, tool-use reliability, multimodal behavior, long-context retrieval, and infrastructure cost on their own data before choosing it for production.
Key Points
- 1Thinking Machines Lab released Inkling on July 15, 2026, with full model weights and an Apache 2.0 license listed in the official model card.
- 2The company reports 975 billion total parameters, 41 billion active per token, multimodal inputs, text output, and a context window of up to one million tokens.
- 3Vendor benchmark charts provide a starting point, but deployment teams still need independent workload testing for quality, latency, memory use, and long-context reliability.
Scoring Rationale
Inkling is the first general-purpose model family from a heavily funded lab led by former OpenAI CTO Mira Murati. Its open weights, Apache 2.0 model-card license, multimodal inputs, and scale make it relevant to teams evaluating customizable models, while vendor benchmarks still require independent validation.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
