Inkling

Thinking Machines Lab released Inkling on July 15, 2026, as its first in-house model and the inaugural entry in a planned model family. The official announcement describes Inkling as a Mixture-of-Experts transformer with 975 billion total parameters and 41 billion active per token, pretrained on 45 trillion tokens span Alongside the main release, the announcement previews Inkling-Small, a lighter-weight variant with 12 billion active parameters trained with a similar recipe, aimed at lower cost and latency. The post also introduces an Inkling Playground in the Tinker console for chatting with the model, and demonstrates the model wri

Thinking Machines Lab released Inkling on July 15, 2026, and the model adopts a Mixture-of-Experts architecture with 975 billion total parameters and 41 billion activated per token, according to a 36Kr report. Pre-training used 45 trillion tokens spanning text, images, audio, and video, and the model supports a maximum The same report notes that Inkling's weights were published under an Apache 2.0 license, enabling self-deployment, and that developers can fine-tune the model through Thinking Machines' Tinker platform. The article also describes the architecture as following the DeepSeek-V3 MoE design, with post-training cold-start da

Thinking Machines Lab released Inkling on July 15, 2026, as an open-weights sparse Mixture-of-Experts transformer with 975 billion total parameters and roughly 41 billion active per token. The model accepts text, images, and audio and supports a 1-million-token context window, shipping under the Apache 2.0 license for direct developer modification. Thinking Machines positions Inkling as a broad, customizable foundation rather than a leaderboard-topping system, pairing downloadable weights with its Tinker post-training platform for fine-tuning workflows. The release reflects a three-part technical bet: large sparse expert capacity without dense-model compute cost per token, native multimodality inside a single decoder instead of bolted-on modality systems, and customization as the core product through Tinker-managed post-training infrastructure. The headline efficiency claim is on Terminal Bench 2.1, where Thinking Machines reports Inkling spends about one-third as many tokens as NVIDIA Nemotron 3 Ultra to match the same coding performance. That token-cost curve matters more than peak scores for long-running agentic workflows.

Sebastian Raschka's architectural breakdown details that Inkling's 66 decoder layers include 2 dense feed-forward blocks at the bottom, followed by layers with 256 routed experts and 2 shared experts, yielding roughly a 4.2% active parameter ratio. The model uses a 5-to-1 local-global attention pattern: 55 sliding-window layers with a 512-token window and 64 query heads, plus 11 global layers. Unusually, Inkling skips RoPE and instead computes a learned, input-dependent relative-position bias from query and key states. Raschka positions Inkling in the same size class as Kimi K2.5 (1T total, 32B active) and GLM-5.2 (744B total, 40B active), noting it is close to GLM-5.2 in active size despite 231B more total parameters. Short convolutions appear inside every decoder block, a detail Raschka notes he has seen less often. Release-time benchmark results from Thinking Machines used an effort setting of 0.99, with architecture diagrams published in the LLM Architecture Gallery.

A detailed third-party breakdown corroborates and expands on the official Inkling launch with operational context for teams evaluating self-hosting. It confirms the core spec sheet — 975B total / 41B active MoE, 45T multimodal training tokens, Apache 2.0 license, full weights on Hugging Face — and notes a practical dis The breakdown also covers the controllable thinking-effort dial and the practical infrastructure considerations of self-hosting a model at this scale, useful for developers and platform teams weighing procurement and compliance. It highlights Tinker as the same-day fine-tuning channel, the self-finetuning demonstration

A Hugging Face community blog post (co-published with Thinking Machines) details day-0 developer tooling for Inkling: BF16 and NVFP4 weight variants, speculative multi-token prediction (MTP) layers for faster inference, and first-day integration in transformers, SGLang, vLLM, and llama.cpp. The post frames Inkling as a The same post was later updated to cover the release of Inkling-Small (276B total / 12B active parameters) and the Inkling-Small-NVFP4 variant, plus MXFP8 weight support, one-click deployment on Hugging Face Inference Endpoints (up to 160 TPS for Inkling-Small), and a real-time voice and image demo. A linked collection

Thinking Machines, the lab founded by former OpenAI CTO Mira Murati, released Inkling on July 15, 2026 as its first open-weights model, allowing developers to download and modify weights directly. Inkling is a Mixture-of-Experts transformer with 975B total and 41B active parameters, pretrained on 45 trillion tokens of text, images, audio, and video, with a 1M-token context window. A smaller sibling, Inkling-Small, was previewed alongside the base release as the first entry in a planned family. The MoE topology largely follows DeepSeek-V3, with 256 routed experts plus 2 shared experts and 6 routed experts active per token, using a sigmoid router. Thinking Machines trained across agentic, reasoning, coding, instruction-following, factuality, vision, and audio tasks to produce a broad base adaptable to fine-tuning, stating plainly that Inkling is not the strongest overall model available today. On Terminal Bench 2.1, Inkling reportedly uses one-third as many tokens as NVIDIA's Nemotron 3 Ultra to reach the same coding performance, making token efficiency the core selling point.

Thinking Machines Lab released Inkling on July 15, 2026 as a 66-layer decoder-only sparse Mixture-of-Experts transformer with 975 billion total parameters and 41 billion active per token. The model is natively multimodal, accepting text, images (40px–4096px), and 16kHz WAV audio up to 20 minutes, producing text outputs. It supports a context window of up to one million tokens, uses BF16/MXFP8/NVFP4 numerics, and is distributed under Apache 2.0 with weights downloadable from Hugging Face. The architecture routes each token to 6 of 256 experts plus 2 shared experts, with a hybrid attention pattern combining local sliding-window and global layers. Images are encoded via a hierarchical patch encoder and audio via discrete token encoding, all projected into a shared hidden space. API access is available through Thinking Machines' Tinker fine-tuning service and third-party inference providers, with distribution aimed at building agentic systems, coding assistants, chatbots, and RAG applications in English and other languages.