Key Takeaways
- Technical debut: Thinking Machines Lab releases Inkling, an open-weight model with 975 billion total parameters, with only 41 billion active during inference thanks to its MoE architecture.
- Competitive standing: 41st place on the Artificial Analysis Intelligence Index, the top-ranked US open-weight model, ahead of Nvidia Nemotron 3 Ultra.
- Operational efficiency: Averages 25,000 tokens in output versus 37,000–43,000 for competing models, released under an Apache 2.0 license with no commercial restrictions.
Market entry with unprecedented scale parameters
Thinking Machines Lab, led by former OpenAI CTO Mira Murati, launched Inkling on July 15, 2026 — the open-weight model with the highest parameter count ever produced on US soil. The figure of 975 billion total parameters places the system in a scale tier that demands dedicated compute infrastructure, yet the significant detail isn't raw size but activation efficiency.
MoE architecture: the relationship between scale and computational cost
Inkling's decoder-only transformer operates across 66 layers with a routing mechanism that selects 6 experts out of 256 per token, supplemented by 2 always-active shared experts. The operational outcome is that the system mobilizes just 41 billion active parameters per request, a sparsity ratio that cuts computational load while preserving the representational capacity of the full-scale model. The context window reaches 1 million tokens, while NVFP4 quantization enables execution on eight Nvidia B300 accelerators or sixteen H200 units, a specification that sets the minimum hardware footprint for deployment.
Multimodal ingestion, single-modality output
Training processed 45 trillion tokens spanning text, images, audio, and video. Images and video are handled via a hierarchical patch encoder, audio through discrete token encoding, with all modalities projected into a shared hidden space processed by the decoder. The architectural constraint requires output to remain text-only: code, structured documents, data. The Apache 2.0 license removes every restriction on downloading, running, and customizing the model, a condition that enables enterprise adoption without license negotiation.
Benchmark metrics: where the model outperforms direct competitors
Ranking 41st on the Artificial Analysis Intelligence Index makes Inkling the highest-scoring US open-weight model, surpassing Nemotron 3 Ultra, which sits at 38 points. On SWE-bench Verified it scores 77.6% against the Nvidia competitor's 71.9%. On VoiceBench it registers 91.4%. On the GDPval-AA v2 index, an Elo of 1238 surpasses both Kimi K2.6 (1190) and DeepSeek v4 Flash (1189), establishing a quantifiable edge across two distinct applied-reasoning benchmarks.
Token efficiency as an operational cost metric
The average consumption of 25,000 output tokens per Intelligence Index task compares against 43,000 for GLM-5.2, 38,000 for Kimi K2.6, and 37,000 for DeepSeek v4 Pro. On Terminal Bench 2.1, Inkling matches Nemotron 3 Ultra while using roughly a third of the tokens, a figure that directly impacts inference overhead at production scale. "Controllable thinking effort" allows tuning the speed/accuracy tradeoff, while an uncertainty-flagging mechanism reduces the incidence of hallucinated output.
Technical debts to external research
Thinking Machines confirmed that Inkling's architecture derives from DeepSeek-V3, and that the post-training phase used data generated by Moonshot AI's Kimi K2.5. This documents the circulation of methodologies across labs regardless of geographic origin, a technical fact the company itself openly states in its release documentation.
Outlook: a customization foundation, not an absolute top-tier model
Thinking Machines' official statement frames the release's ambitions narrowly: "Inkling isn't the overall most capable model available today, open or closed. It is instead an excellent open-weight foundation for customization." The strategic positioning doesn't aim for absolute leaderboard supremacy, but rather the creation of a modular, permissively licensed base infrastructure. The immediate impact on the US open-weight market is the emergence of a technical standard with a better efficiency-to-performance ratio than anything previously available, a condition that redefines the competitive benchmark for the sector's upcoming releases.
