Cognition has officially released Inkling, a 975-billion-parameter open-weights model family engineered for broad customization and multimodal reasoning. Built as a Mixture-of-Experts Transformer, the flagship variant activates 41 billion parameters per token and accommodates context windows up to one million tokens. Pretrained from scratch on 45 trillion tokens encompassing text, images, audio, and video, Inkling is positioned as a flexible foundational base rather than a narrowly optimized frontier system, prioritizing adaptability, cost efficiency, and native cross-modal processing. A central architectural innovation is Inkling controllable thinking effort. Developers can dynamically adjust reasoning depth to balance computational expense against task performance. Benchmark evaluations across agentic coding, mathematical reasoning, and instruction following demonstrate that Inkling achieves parity with higher-parameter competitors while consuming approximately one-third the tokens. The model utilizes a sigmoid-based router, interleaved sliding-window and global attention layers, and relative positional embeddings optimized for long-context extrapolation. Audio and vision inputs bypass traditional encoders, employing discrete spectral and patch-based embeddings that integrate seamlessly with textual tokenization for unified reasoning. Safety and epistemic reliability were prioritized during post-training. The development team applied reinforcement learning against proper scoring rules to calibrate confidence metrics, encouraging factual hedging and abstention when uncertain. Instruction-following and accuracy were reinforced through dual automated graders that penalize hallucinations without degrading helpfulness. Independent safety audits confirmed that Inkling maintains strong refusal rates against harmful requests while minimizing over-refusal on benign analogs, a critical equilibrium for open-weight deployment. To accelerate developer adoption, Cognition is opening Inkling for fine-tuning on its Tinker platform, accompanied by a dedicated playground interface for interactive benchmarking. The model is accessible via APIs from Together, Fireworks, Modal, and Databricks, alongside open-source integrations in vLLM, SGLang, and llama.cpp. Full checkpoints and a compressed NVFP4 variant optimized for NVIDIA Blackwell architecture are available on Hugging Face. Tinker users can currently access the model at a temporary 50 percent discount. Alongside the flagship release, Cognition shared a preview of Inkling-Small, a 276-billion-parameter variant with 12 billion active parameters. This lightweight counterpart matches or exceeds the larger model on multiple reasoning and agentic benchmarks while delivering substantially lower latency and inference costs, making it suitable for high-throughput workflows such as automated evaluation and synthetic data generation. Full weights for Inkling-Small will follow final validation, with Cognition indicating continued iteration across the family to expand compute scaling and training pipeline efficiency.