Inkling: The Biggest Open Multimodal Model Yet — 975B Params, 1M Context, Day-0 in vLLM and GGUF

TL;DR: Inkling is a 975B-param open-weight MoE model (41B active per token) from Thinking Machines that natively processes text, images, and audio in a 1M-token context window — available today on HuggingFace with day-0 framework support.

Thinking Machines released Inkling on July 15 — the largest open multimodal model to date, and the first to natively accept all three modalities (text, images, audio) without adapter bolts. The key numbers:

• 975B total parameters, 41B active via Mixture-of-Experts routing — only 4% of parameters fire per token, making inference tractable
• 1M token context window — trained on 45 trillion tokens of text, images, audio, and video
• Two official variants: full BF16 (requires 2TB VRAM on Hopper) and NVFP4 (600GB on Blackwell); community contributors already have 1-bit GGUF quantizations that cut memory ~95%
• Day-0 framework support: transformers v5.14.0 (released same day), vLLM, SGLang, llama.cpp, and Unsloth

Why it matters: Until now, natively multimodal open models topped out well below 100B. Inkling changes the calculus for teams that need self-hosted control over audio-reasoning, document-analysis, or agentic pipelines — the model's MoE architecture and aggressive GGUF compression mean server-grade hardware (rather than data-center scale) can run it. Inference Providers on HuggingFace offer cloud access if you don't have the hardware yet.

Sources: Inkling announcement — HuggingFace Blog · Inkling model card — HuggingFace · Transformers v5.14.0