← Back to feed

Thinking Machines Lab, the AI company founded by former OpenAI CTO Mira Murati in late 2024, has released its first model. Inkling is a 975B-parameter mixture-of-experts transformer with 41B active parameters, a 1M token context window, and native support for text, image, and audio inputs. Full weights are available today.

The release is notable less for raw benchmark performance — Thinking Machines is explicit that Inkling is “not the strongest overall model available today, open or closed” — and more for what it is: a credible open-weights multimodal base from a lab that spent nearly two years building from scratch with one of the field’s most recognizable leaders.

Architecture

Inkling makes several departures from standard transformer practice. Rather than RoPE for positional encoding, it uses relative attention: each layer learns position directly in the attention logits through a per-token, per-head relative feature. Decoder layers alternate between global attention and sliding window attention in a 5:1 sliding-to-global ratio, with a short 1D convolution over hidden states handling local representations.

The MoE layer has 256 experts. Each token routes through 6 of them plus 2 always-active shared experts — a design Thinking Machines calls a shared-experts sink. With 41B active parameters, inference costs track closer to a mid-range dense model despite the 975B total count.

For multimodal input, vision and audio pass through lightweight tower modules — a hierarchical MLP patchifier for images and a discretized mel spectrogram for audio — rather than separate heavyweight encoders. Text, images, and audio were all present in pretraining: 45 trillion tokens total.

Deployment Footprint

The full BF16 checkpoint requires at least 2TB of aggregate GPU memory. An NVFP4 quantized checkpoint lowers that to 600 GB. Day-0 support is available in Transformers, SGLang, and llama.cpp. Modal reports 250 tokens per second per user on 8x B200s at 2.5M TPM, 67% faster than the model’s built-in speculative path at matched throughput.

Alongside the full model, Thinking Machines is releasing Inkling-Small, a lighter variant with 12B active parameters built on the same recipe.

Positioning

The model is available for fine-tuning on Tinker, Thinking Machines’ training platform, with an Inkling Playground developer console. The company frames Inkling as a base for customization rather than a frontier benchmark challenger — a deliberate positioning that sets it apart from labs racing to claim SOTA on standard eval suites.

As a demonstration, Thinking Machines had Inkling fine-tune itself: the model wrote its own fine-tuning job, ran it, and evaluated the result on Tinker.

The Inkling-Small model and the Tinker fine-tuning infrastructure point toward the longer play: a platform for domain adaptation at scale, with Murati’s lab owning the full model-to-deployment stack. Inkling is available on Hugging Face, Baseten, and Modal.