Open-weights multimodal MoE model (276B total, 12B active) from Thinking Machines; text, image and audio in, text out.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
Inkling Small is a smaller open-weights variant of Thinking Machines Lab's Inkling model family. It is a natively multimodal sparse Mixture-of-Experts transformer with 276B total and 12B active parameters: 42 layers, 6 of 256 experts routed per token plus 2 shared experts, and hybrid local/global attention. Images and audio are projected into a shared hidden space and processed jointly with text by the decoder.
The model accepts text, images (roughly 40–4096 px per side) and 16 kHz WAV audio (ideally under 2 minutes), and generates text only. It is aimed at developers building agentic and tool-use systems, coding assistants, chatbots and RAG applications, and is released for research, fine-tuning and integration. BF16 and NVFP4 weights are published, and local serving recipes exist for SGLang, vLLM, Unsloth, TokenSpeed and Hugging Face Transformers.
It can be tried in the Tinker Playground or accessed through the Tinker API and third-party inference providers. Creator reviews describe it as behind DeepSeek V4 Flash, GLM 5.2 and Kimi K3 in general intelligence on Artificial Analysis and slower, but stronger on audio and vision. The BF16 weights are reported at about 532 GB. Use is also subject to an acceptable use policy.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Smaller open-source omnimodal model (text, audio, image, video) with 276B total and 12B active parameters, about a quarter the size of the full Inkling. It trails DeepSeek V4 Flash in intelligence but is stronger on audio. Weights are 532 GB.
Open-weights multimodal MoE of almost 1T parameters (41B active) with 1M context, trained on 45T tokens; also a 276B Inkling small preview. Creator found it behind GLM 5.2 and Kimi K3 on Artificial Analysis and slower, but liked its audio and vision.
Closest neighbours by meaning, not just by category.
Model · Thinking Machines
Open-weights multimodal MoE model (594B total, 17B active) from Thinking Machines, available via Tinker.
Model · Thinking Machines Lab
Open-weight 952B-parameter language model from Thinking Machines Lab, its first public open-weight release.
Model · MiniMax
Open-weights native multimodal MoE model from MiniMax, ~428B parameters (~23B active) with 1M-token context.
Model · Meituan
Meituan's open-weights 1.6T-parameter multimodal MoE model (48B active) for agentic coding and visual tasks.
Model · OpenBMB
Small open-weights multimodal model from OpenBMB for image and video understanding on phones and edge devices.
Model · Accio Lab
Open-weights 35B MoE (3B active) agentic model from Accio Lab, post-trained from Qwen3.6-35B-A3B.