Thinking Machines drops Inkling: a 975B-param open model that sees, hears, and reads

Forget the single-mode giants. Thinking Machines just put Inkling on Hugging Face, and it’s not just another large language model — it’s a 975-billion-parameter, Mixture-of-Experts beast built to reason across text, images, and audio natively, all within a massive 1 million token context window. Trained on a staggering 45 trillion tokens of multimedia data, Inkling is the first open model of this scale to go truly multimodal without bolting on separate encoders as an afterthought.

The architecture is dense with clever engineering. Instead of the ubiquitous RoPE positional encoding, it uses relative attention, learning position relationships directly in the attention logits. It alternates between global and sliding window attention in a 5:1 pattern to handle that 1M context efficiently. A unique short convolution layer over hidden states handles local features, while its MoE router juggles 256 total experts, activating 6 routed and 2 shared experts for any given token. The vision and audio towers are refreshingly simple — a hierarchical MLP patchifier for pixels and a discretized mel spectrogram for sound — but don’t underestimate them. The image tower even includes a temporal dimension for video, though Thinking Machines is upfront that out-of-the-box video performance hasn’t been evaluated yet.

Of course, you need serious hardware. The full BF16 checkpoint demands 2TB of VRAM. But the team also released a well-calibrated NVFP4 variant that shrinks the footprint to 600GB, and it’s already got day-zero support in transformers, SGLang, and vLLM. You can even run it locally via llama.cpp with GGML quants. “We’ve tinkered with this model to build some demos and explore the architecture,” the team writes, “and we think it’s great for building a new wave of multimodal reasoning apps.” That’s the key phrase: domain adaptation. This isn’t a polished consumer product; it’s a powerful, openly-available foundation for teams that need to fine-tune a model that thinks across senses.

I’m genuinely curious to see what builders do with the audio modality in particular. Most multimodal work still treats sound as an afterthought. Having speech, environmental audio, and music as a first-class input alongside text and images — inside a model that can reason over a novel-length context — opens up applications in accessibility, media analysis, and robotics that feel underexplored. The hardware barrier is real, but the NVFP4 quant is a pragmatic olive branch to the community. If the fine-tuned derivatives deliver, Inkling could define its own category.