Hugging Face BlogJul 15, 2026, 08:00 UTC+8

Thinking Machines releases Inkling: a 1T-param, natively multimodal open model

Welcome Inkling by Thinking Machines

Inkling is an open ~1T-param model that natively accepts image, audio, and text inputs with a 1M context window. Trained on 45T multimodal tokens, it uses a MoE architecture with 975B total and 41B active parameters. It includes MTP speculative decoding layers for faster inference and ships in BF16 and NVFP4 variants. Hugging Face provides day-0 support in transformers, SGLang, vLLM, and llama.cpp, covering agentic coding, multimodal vision, and audio tasks.

Why it matters: A new player, Thinking Machines, open-sources a trillion-parameter multimodal MoE model with solid specs (975B/41B activated, 1M context, 45T tokens trained) and MTP speculative decoding. H and K both hit, but R is weak — the team has no name recognition, no emotional anchor. ...

Read the original ↗Export Markdown

Take

The reason to click: 975B total params but only 41B active per token — that's a MoE setup designed to keep inference cheaper while natively handling image, audio, and text, not bolting on encoders after the fact. 1M context window, 45T multimodal training tokens, and built-in multi-token prediction layers for speculative decoding. Hugging Face shipped day-0 support across transformers, vLLM, SGLang, and llama.cpp, which tells you the community was ready for this one.

Where I'd hold back: the post focuses on architecture and integration, not benchmarks. No concrete training cost or inference latency numbers yet. 41B active is lighter than 975B, sure, but it's still not a single-GPU casual load — the NVFP4 compressed variant should help, but we need real-world numbers. For now, treat it as an open multimodal foundation model, not a closed-source competitor.

Source: Hugging Face Blog · huggingface.co
Published: Jul 15, 2026, 08:00 UTC+8 · Jul 15, 2026