Thinking Machines releases Inkling, a 975B MoE model with Apache 2.0 open weights
The company says only ~41B parameters are active per token and publishes 2TB-to-600GB memory footprints for deployment.
Thinking Machines has released Inkling, a from-scratch Mixture-of-Experts model it says totals 975 billion parameters while activating roughly 41 billion per token. The weights are described as available on Hugging Face under an Apache 2.0 license, alongside unusually explicit GPU memory requirements that map the model’s “openness” to real deployment constraints.
Key Takeaways
- Thinking Machines released Inkling on July 15, 2026, positioning it as the company’s first model trained from scratch.
- Inkling is described as a 975B-parameter Mixture-of-Experts model with roughly 41B parameters active per token, using sparse routing rather than dense compute.
- The architecture is specified as 66 layers with 256 experts per layer, routing six experts per token plus two shared experts that run on every token.
- Deployment footprints are stated directly: about 2TB of combined GPU memory for full precision and around 600GB for a quantized checkpoint.
Inkling lands as an Apache 2.0, from-scratch MoE release
Thinking Machines released a model called Inkling on July 15, 2026. The company is described as making the weights available on Hugging Face under an Apache 2.0 license, which typically permits broad reuse and modification subject to the license terms.
Inkling is framed as a Mixture-of-Experts (MoE) system, meaning the model stores many parameter “experts” but only runs a subset for each generated token. The headline sizing in the technical breakdown is the gap between a 975 billion total parameter count and roughly 41 billion parameters involved in processing any single token, or about 4% active at a time.
The packet’s details come from a secondary technical breakdown that explicitly warns it is compiled from “publicly shared details from various sources” and asks readers to flag inaccuracies. The underlying references and primary release artifacts are not included here, so the mechanism is clear but direct verification is limited to what’s in the excerpt.
The numbers traders will anchor on: 2TB full precision, ~600GB quantized
Inkling’s release reads like an infrastructure story because it attaches concrete memory footprints to the model checkpoints, not just parameter counts. The breakdown states the full-precision checkpoint needs at least 2 TB of combined GPU memory, given as eight NVIDIA B300 cards or sixteen H200 cards. It also describes a quantized checkpoint at around 600 GB that fits on four B300 cards.
Those figures matter because they turn “open weights” into a capacity planning problem. If a model can be downloaded and run, traders can map that claim to a rough hardware bill of materials, then to hosted inference economics, then to demand for GPU capacity and the stacks that serve it. A 975B headline number can be narrative fuel on its own, but a 2TB requirement is a constraint that shows up in procurement, colo planning, and pricing.
MoE is the bridge between those two realities. Inkling’s design is described as separating storage cost from per-token inference cost: you still have to load the full checkpoint into memory to run it, but each token only activates a small fraction of the parameters. That is the pitch for why a model can be “very large” on paper while being cheaper to run per token than its total parameter count implies.
How Inkling gets to 1M tokens: 55 sliding-window layers and 11 full-attention layers
Inkling is described as supporting a one million token context window, which is the maximum amount of text it can consider at once. The mechanism offered for making that feasible is attention sparsity: most layers use sliding-window attention, where each token only attends to a limited recent range, and a smaller number of layers use full attention, where tokens can attend to the entire prior sequence.
The breakdown describes a 5:1 alternation between sliding-window layers and full-attention layers. It also cites vLLM integration notes that put numbers on the split: 55 sliding-window layers and 11 full-attention layers out of 66 total layers.
This is a specific architectural tradeoff, not a marketing claim in isolation. Sliding-window layers keep compute from exploding with sequence length, but they also risk losing long-range recall because distant facts are not directly visible in most layers. The periodic full-attention layers are the escape hatch, providing global “refresh points” where information from far earlier in the sequence can be pulled forward and then carried locally again.
What to verify before markets price it in: sources, truncation, and missing benchmarks
The gating issue for treating Inkling as a catalyst is verification of the primary artifacts. The packet does not include a Thinking Machines model card, a direct Hugging Face link, or the referenced documents behind the excerpt’s bracketed citations, so the first confirmation point is the actual model page and the exact Apache 2.0 licensing and usage terms attached to the weights.
The second missing piece is performance evidence. No benchmarks, latency numbers, throughput claims, or long-context evaluations are provided in the packet, and the excerpt is truncated mid-discussion of position encoding. Third-party evals, especially tests that probe long-context behavior near the one million token window, would be the fastest way to separate “supports 1M tokens” from “accepts 1M tokens but forgets what matters.”
The third confirmation is deployability at the stated footprints. The breakdown gives specific memory targets for full precision and quantized checkpoints, and it references vLLM integration notes for the 55/11 attention split. Reports of successful runs on those configurations, plus evidence that common inference stacks handle the attention pattern as described, would tighten confidence that this is a model people can actually host rather than just download.
My read: open weights + explicit footprints tighten the feedback loop between model launches and compute trades
The part that decides whether Inkling matters to markets is not the 975B headline, it is the combination of sparse activation and explicit deployment footprints. A MoE model that only activates ~41B parameters per token can plausibly change the per-token cost curve without giving up the marketing aura of “nearly a trillion parameters,” and that is exactly the kind of ambiguity traders usually have to guess at.
The threshold that matters is whether the Apache 2.0 Hugging Face listing and the 2TB-to-600GB memory claims hold up in real deployments, then get echoed by fine-tunes, quantization variants, and hosted inference offerings. If those confirmations arrive, “open long-context” stops being a narrative and starts being a measurable GPU-capacity input that can propagate through the stack.
