Inkling

AIME 2026: 97.1%, HLE w/ Tools: 53.2%, GPQA Diamond: ~90%, competitive with Nemotron 3 Ultra, GLM 5.2, and DeepSeek V4 Pro across text/agentic/multimodal evaluations. Artificial Analysis Intelligence Index: 41 | Broad-domain generalist work across text, image, and audio inputs — domain adaptation via fine-tuning on Tinker, agentic coding and tool use, multimodal reasoning (charts, documents, audio transcription/analysis), instruction following, forecasting with calibrated confidence, and controllable-effort reasoning (adjust thinking time to balance speed vs depth). First model in a planned family (Inkling-Small with 276B/12B-active in preview). Apache 2.0 licensed open weights on Hugging Face. **Caveats: Thinking Machines explicitly states Inkling is "not the strongest overall model available today, open or closed" — it is designed as a balanced, customizable base rather than a benchmark topper. AA Intelligence Index of 41 trails frontier proprietary models (GPT-5.6 Sol ~59, Opus 4.8 ~57) and leading open-weights peers (Nemotron 3 Ultra 48, GLM-5.2 ~52). Post-training data included outputs from other open-weight models (Kimi K2.5) — fully self-contained post-training planned for next generation. Very new (Jul 15, 2026) with limited independent benchmark verification.**