Thinking Machines drops Inkling, a 975B open-weights multimodal MoE

Inkling is a 975B-parameter open-weights Mixture-of-Experts model from Thinking Machines that activates 41B parameters per token, supports a 1M context window, and natively reasons across text, audio, and vision.

Its standout operational feature is controllable thinking effort, letting you trade latency and token cost directly against output quality.

What it does

Inkling is built from scratch as an encoder-free multimodal foundation rather than a text model with vision adapters bolted on.

Audio feeds directly into the network as dMel spectrograms, while images are processed as 40x40 patches through a four-layer hMLP. In inference workflows, it can execute Python code to zoom, crop, and inspect images iteratively, though its final output modalities remain restricted to text, code, and structured artifacts.

On agentic tasks, it can generate full-stack web apps in a single pass or iterate on complex codebases across dozens of review loops. By tuning its thinking effort slider, it matches the coding performance of Nemotron 3 Ultra on Terminal Bench while consuming just one-third of the tokens.

Thinking Machines also previewed Inkling-Small, a lighter 276B variant with 12B active parameters for lower-latency deployments.

Hardware specs and where to run it

Full-scale Inkling is heavy iron that will not fit into consumer workstations.

The unquantized weights on Hugging Face weigh in at 1.9 TB, while an NVFP4 quantized checkpoint requires 592 GB. Community conversions from Unsloth provide GGUF builds down to a 1-bit quantization at 270 GB, though aggressive quantization takes a heavy bite out of model intelligence.

Developers looking to test without provisioning dedicated clusters can chat with the checkpoint directly in the Tinker Playground or run custom fine-tuning jobs on the Tinker platform.

Why it matters

If your pipeline requires native voice and vision reasoning inside an open-weights stack you can modify freely, Inkling is one of the few viable bases at this scale.

Just keep expectations calibrated on pure text and code speed: third-party evaluations show it lagging behind GLM 5.2 and DeepSeek V4. We see Inkling less as an out-of-the-box chatbot and more as a heavy-duty starting point for teams with the compute to fine-tune specialized internal agents.