Inkling

Inkling is Thinking Machines Lab's (TML) open-weights foundation model announced on July 15, 2026. It is the first model trained from scratch by the laboratory founded by Mira Murati (former CTO of OpenAI). The architecture is a Mixture-of-Experts Transformer with 975B total parameters, 41B active per token, and a context window of up to 1M tokens. Pretraining used 45 trillion tokens spanning text, images, audio, and video.

The model supports controllable thinking effort — developers can slide the trade-off between quality and generated token count. TML shows Inkling matching Nemotron 3 Ultra on Terminal Bench 2.1 at roughly one third of the tokens. Native multimodal inputs cover text, images (40×40 patches through a four-layer hMLP) and audio (discrete dMel spectrograms). Outputs are text and code.

The MoE architecture broadly follows DeepSeek-V3: each MoE layer has 256 routed experts plus 2 shared experts, with 6 routed experts active per token, a sigmoid router with auxiliary-loss-free load-balancing bias. Attention interleaves sliding-window and global layers at a 5:1 ratio with 8 KV heads, input-dependent relative positional bias (not RoPE), and short convolutions after key/value projections and the residual branches. Post-training bootstrapped with SFT on synthetic data from open-weights models including Kimi K2.5, followed by large-scale asynchronous RL over 30M+ rollouts. Trained on NVIDIA GB300 NVL72 systems. Hybrid optimizer — Muon for large matrix weights, Adam for the rest.

Benchmarks (at effort 0.99): HLE 29.7% (text) and 46.0% (with tools), AIME 2026 97.1%, GPQA Diamond 87.2%, SWE-Bench Verified 77.6%, SWE-Bench Pro Public 54.3%, Terminal Bench 2.1 63.8%, GDPVal-AA v2 1238 Elo, MCP Atlas 74.1%, Tau 3 Banking 23.7%, BrowseComp 77.1%, SimpleQA Verified 43.9%, IFBench 79.8%, Global-MMLU-Lite 88.7%, MMMU Pro 73.5%, Audio MC 56.6%, MMAU 77.2%, VoiceBench 91.4%, FORTRESS Adversarial 78.0%, StrongREJECT 98.6%. On Design Arena Agentic Web Dev (blinded human eval) Inkling scores 1257 — among the strongest open-weights models.

Availability — full weights on Hugging Face (original and NVFP4 checkpoint for NVIDIA Blackwell) under repo thinkingmachines/inkling. Fine-tuning on TML's Tinker platform (context length options 64K and 256K). API inference through partners: Together, Fireworks, Modal, Databricks, Baseten. Open-source support in SGLang, Miles, vLLM, TokenSpeed, llama.cpp, and Hugging Face Transformers. TML also previewed Inkling-Small — a smaller MoE variant with 276B total / 12B active parameters that matches its larger sibling on many benchmarks thanks to improved pretraining data and recipe.