Thinking Machines Lab, founded in early 2025 by former OpenAI CTO Mira Murati, released Inkling — a frontier open-weights model weighing 975 billion parameters — on Wednesday (July 15–16, 2026 coverage window).
Native 16-bit inference requires more than two terabytes of GPU memory (roughly eight Nvidia B300 accelerators, or sixteen H200s). An NVFP4 quantized version runs on about half the GPUs. The Register calls it the largest American open-weights model to date, comparable in size to Chinese models such as DeepSeek V4, GLM 5.2, and Kimi K2.6. Benchmarks show it competitive with those open models while trailing proprietary systems from Anthropic and OpenAI.
Inkling uses a Mixture-of-Experts architecture inspired by DeepSeek-V3: 256 routed experts plus two shared experts; six experts activate per token (~41 billion active parameters). It was trained from scratch on Nvidia GB300 NVL72 systems using 45 trillion tokens of text, images, audio, and video. Features include a 1-million-token context window, Apache 2.0 licensing, and “reasoning”/chain-of-thought behavior with tunable thinking-token efficiency. Thinking Machines claims it matches Nvidia Nemotron 3 Ultra on Terminal Bench 2.1 using roughly one-third the tokens.
Availability: Tinker platform (API, customization, fine-tuning); planned third-party APIs (TogetherAI, Fireworks, Modal, Databricks, Baseten); Hugging Face downloads with support for vLLM, SGLang, Miles, TokenSpeed, and Llama.cpp. Inkling-Small (276B total / 12B active MoE) is in preview; full weights planned after testing.
Supplementary detail from THE DECODER (Jul 16, 2026): multimodal (text, images, audio); Artificial Analysis Intelligence Index score 41 (top U.S. open-weights); strong agentic GDPval-AA v2 Elo 1,238; weak factual accuracy (40% accuracy, 63% hallucination rate on AA Omniscience); pricing around 4.68 per million input/output tokens at 64K context.