News · 2026-07-30

Thinking Machines ships Inkling-Small's open weights - all 532 gigabytes of them

Thinking Machines has released the open weights for Inkling-Small, a 276-billion-parameter mixture-of-experts model that activates 12 billion parameters per token and takes text, images and audio as input. The download is roughly 532 gigabytes across 33 files, published under Apache 2.0 with a separate acceptable-use policy attached. The company had promised in mid-July to "release its full weights once that work is complete"; this is that release.

Key facts

• 276B total parameters, 12B active - a 42-layer decoder routing each token to 6 of 256 experts, plus 2 shared experts always on.
• ~532 GB of BF16 weights across 33 files on the Hugging Face repository, created July 27 and actively updated through July 30.
• Natively multimodal: text, image and audio in, text out, with images encoded by a hierarchical patch encoder and audio by discrete tokens.
• Licence: Apache 2.0, plus a separate Model Acceptable Use Policy.

The architecture is the now-standard answer to a hard problem: how do you get the knowledge capacity of a very large model without paying to run all of it on every word? A mixture of experts splits the network's feed-forward layers into many specialised blocks and sends each token to only a few. Inkling-Small holds 256 experts per layer and routes each token to six of them, plus two shared experts that see everything. The result is a model that knows like a 276-billion-parameter system and computes like a 12-billion one.

The honesty of that framing matters, because the memory bill does not shrink. All 256 experts have to be somewhere the model can reach quickly, which is why the download is over half a terabyte. This is the same trap that shows up every time a large open-weight model ships: Kimi K3's weights were downloadable in July but the practical floor to run them was eight datacenter GPUs. Sparse activation buys speed, not a smaller disk.

What closes that gap in practice is quantization, and the ecosystem moved fast. Alongside the official NVFP4 build from Thinking Machines, the Hugging Face quantizations page already lists around nineteen derived repositories, including a GGUF build from Unsloth and MLX conversions at four, six and eight bits for Apple Silicon. The model card itself points to deployment recipes for SGLang, vLLM, Unsloth and Hugging Face. Two days after the weights appeared, the paths to running them locally were already paved.

The licence deserves a closer look than it usually gets. The repository metadata says Apache 2.0, which is about as permissive as software licensing gets. But Thinking Machines also publishes a Model Acceptable Use Policy that users agree to on accessing, downloading or using the model materials, with explicit restrictions covering illegal activity, child safety, violence, harassment and discrimination, privacy, deception and misinformation, and several other categories. Whether that combination is enforceable in the way a licence is remains an open question the industry has not settled. What is clear is that "Apache 2.0" on the repository badge is not the whole agreement, and the phrase "open weight" continues to describe a spectrum rather than a state.

The release lands in the middle of an unusually crowded week for open weights - LG shipped K-EXAONE 2.0 at 750 billion parameters under a genuinely unrestricted Apache licence a day earlier, and the broader shift in who serves what has been visible for months in Chinese open models passing US models in OpenRouter token share. Thinking Machines' pitch, endorsed by the independent evaluator Artificial Analysis in July, is that Inkling is the leading US open-weights model. That is a claim about a national field as much as a technical one.

The caveat is about what the weights do not come with. The model card does not state a context window for Inkling-Small, and Hugging Face's own metadata reports the size as 266 billion parameters where the card body says 276 billion total - a small discrepancy, but one worth knowing before quoting a figure. More consequentially, a claim circulating on release day held that Inkling-Small had topped ARC Prize's open-weight leaderboard. It has not; the ARC Prize result page belongs to the larger Inkling model, and is dated July 17.

Primary source, verified: read the paper →

Ground Truth. (2026, July 30). Thinking Machines ships Inkling-Small's open weights - all 532 gigabytes of them. Ground Truth. https://groundtruth.day/news/thinking-machines-ships-inkling-small-open-weights-532-gigabytes-of-them.html

@misc{groundtruth:thinking-machines-ships-inkling-small-open-weights-532-gigabytes-of-them,
  title  = {Thinking Machines ships Inkling-Small's open weights - all 532 gigabytes of them},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {jul},
  url    = {https://groundtruth.day/news/thinking-machines-ships-inkling-small-open-weights-532-gigabytes-of-them.html}
}

Topics: open-weights · model-release · mixture-of-experts · multimodal · thinking-machines · licensing