What is Thinking Machines Inkling? 975B Open Weights, Multimodal Architecture, and Technical Analysis

Thinking Machines Lab Launches Inkling Model: 975B Open Weights and Multimodal Technology Analysis

Thinking Machines Lab has finally unveiled their first product. Led by former OpenAI CTO Mira Murati, this startup released Inkling, a multimodal model trained from scratch, on July 15, 2026. This time, they used an Apache-2.0 license to publicly release the 975B parameter model weights. In a market currently dominated by high-end commercial closed-source models, releasing weights of such a large-scale model still holds significant value for the research community and enterprise deployment.

Interestingly, they were very frank in their official announcement: Inkling is not the strongest model currently on the market. Its positioning is more like a highly balanced base model that is convenient for fine-tuning and customization.

975B Parameters and Mixture-of-Experts (MoE) Design

Inkling has a 66-layer decoder architecture with a total of 975B parameters. However, because it uses a Mixture-of-Experts (MoE) architecture, only 41B parameters are activated per inference, which keeps the computational costs from being excessively high.

Specifically, its routing mechanism selects the 6 most relevant experts from 256 expert modules, combined with 2 resident shared experts to handle tasks. The routing part uses a Sigmoid-based router and auxiliary-loss-free load balancing bias, mainly to solve the common expert load imbalance problem in MoE. Regarding the attention mechanism, it alternates local attention and global attention in a 5:1 ratio, combined with relative position embeddings to handle long texts.

Native Multimodal Architecture with No External Encoder (encoder-free)

Compared to splicing-style multimodal models that rely on external encoders, Inkling adopts a native encoder-free multimodal architecture. It was pre-trained on 45 trillion (45T) tokens including text, images, audio, and video.

In visual processing, it uses a four-layer hMLP structure to convert images into 40x40 pixel patch encodings; for audio, it uses dMel spectrogram technology to discretize 16kHz sound signals into tokens. Because all inputs are jointly processed in the same latent space, combined with a 1-million (1M) token ultra-long context window, Inkling performs quite fluently when understanding multimodal context.

Telegraphic Chain of Thought and Controllable Thinking Quota

Official technical reports indicate that during the reinforcement learning (RL) training process of over 30 million (30M) rollouts, the research team discovered an interesting phenomenon: Inkling’s chain-of-thought (CoT) self-evolves. Initially, it would organize its thoughts using complete grammatical structures like “We need to understand…”; but as training progressed, it stripped away unnecessary articles and conjunctions, abbreviating them into telegraphic “We need determine.” This approach does not affect the output results but saves a lot of tokens.

Coupled with this characteristic, Inkling provides a “controllable thinking effort” parameter, ranging from 0.2 to 0.99. Developers can manually adjust the “thinking degree” of the model to control cost and precision. For example, according to benchmarks showcased officially, when achieving coding performance equivalent to Nemotron 3 Ultra in Terminal Bench 2.1 tests, Inkling only consumed about one-third of the tokens compared to the other. If you are interested in this evolution of chain-of-thought, you can look at their Inkling model card.

Hardware Deployment and Tinker Fine-Tuning Platform

Getting a 975B model up and running is no easy task. If using the full BF16 format, you need approximately 2TB of VRAM, which basically requires 8 NVIDIA B300s or 16 H200s. However, on the Blackwell architecture, you can use the optimized NVFP4 format to reduce VRAM requirements to 600GB.

If hardware resources are limited, they also offer a smaller “Inkling-Small” preview version. Total parameters are 276B, only 12B parameters are activated each time, and the runtime threshold is much lower, making it very suitable for edge deployment or proof-of-concept (PoC). In addition, combined with their Tinker custom fine-tuning platform, the open-source version provides two context lengths: 64K and 256K. Now, developers can download these weights from the Inkling page on Hugging Face and execute inference using the tml-renderers package.

Epistemology and Safety Defenses

In terms of safety, Inkling performs well in epistemological assessments; simply put, it knows its own knowledge blind spots. When asked about uncertain areas, it tends to give conservative, qualified guesses rather than confidently spouting nonsense.

In addition, while blocking malicious requests, the model avoids over-defensiveness as much as possible. In FORTRESS safety assessments, it can accurately refuse harmful instructions such as weapons or violence, but the false positive rate for those “seemingly sensitive but actually harmless” gray-area questions is quite low. According to data from the official model card, it scored 98.6% in StrongREJECT’s interception tests for harmful requests.

Inkling may not be the top-tier benchmark monster, but with its open-source weights, native multimodal design, and telegraphic thinking efficiency, it is currently a new option worth paying close attention to for developers who wish to deploy and fine-tune models themselves or reduce dependence on closed-source APIs.