Inkling

Inkling is a Mixture of Experts (MoE) large language model with 975B total parameters but only 41B active per token developed by Thinking Machines Lab. Released in July 2026, it supports text and vision and audio inputs with a 1M context window, released under the Apache 2.0 + AUP license, allowing commercial use. General-purpose native multimodal model with 256 routed experts, 6 selected plus 2 shared experts, hybrid local/global attention, image input, and 16 kHz audio input up to about 20 minutes. Thinking Machines reports GPQA-Diamond 87.2, SWE-Bench Verified 77.6, and Terminal-Bench 2.1 63.8. BF16 needs at least 2 TB aggregate VRAM; the official NVFP4 checkpoint still needs roughly 600 GB.

To run Inkling locally, you need approximately 667.5 GB of VRAM at Q4_K_M quantization with 8k context. 1 of the GPUs we track can run it fully in VRAM, with a further 0 able to offload to system RAM. At Q4_K_M it requires 667.5 GB — more than any single consumer GPU. A multi-GPU server or 80 GB datacenter GPU is required. At Q8_K_M (1163.3 GB), you get near-FP16 quality while still fitting on large server or multi-GPU setups. FP16 requires 2186.5 GB, limiting it to datacenter-class hardware with 80 GB+ VRAM. As a MoE model, inference speed depends on active parameters (41B) rather than total size, so once it fits in VRAM it runs noticeably faster than dense models of the same parameter count.

Inkling is designed for reasoning and problem-solving. Thanks to its MoE architecture with only 41B active parameters, it delivers strong quality while keeping inference latency manageable for local use.