The interesting thing about Inkling is not where it lands on a leaderboard — it is that a US lab put a frontier-scale, genuinely Apache-2.0 model on Hugging Face. For anyone deciding what they can legally and practically self-host, the license and the provenance move more than another benchmark point ever could.
On 15 July, Thinking Machines Lab — the outfit Mira Murati founded in early 2025 after leaving OpenAI as CTO — released Inkling, its first model, and released it open-weight-first. Inkling is a 975-billion-parameter mixture-of-experts model with about 41 billion parameters active per token (six experts fire on any given token), trained from scratch on 45 trillion tokens of text, images, audio, and video, with a one-million-token context window. A smaller Inkling-Small (roughly 276B, 12B active) ships alongside it.
The distribution is the part that signals intent. The weights are on Hugging Face under Apache 2.0; day-one hosting spans Together, Fireworks, Modal, Databricks, and Baseten; and runtime support covers vLLM, SGLang, and llama.cpp out of the gate. Fine-tuning runs through the lab's Tinker platform. This is not a research checkpoint tossed over the wall — it is a model packaged to be deployed.
Plenty of models call themselves open; far fewer are Apache 2.0. Much of what Western labs ship as "open weight" arrives wrapped in a custom community or acceptable-use license that caps commercial scale, restricts redistribution, or fences off fine-tuning — terms your legal team has to read before you build anything real on them. Inkling ships with none of that: download, modify, integrate, and commercialize, royalty-free. For a frontier-scale model that is genuinely unusual, and it is the detail that changes the calculus rather than the leaderboard.
Provenance is the quieter half of the same point. Until now the most capable permissively-licensed open weights came largely from Chinese labs — DeepSeek, GLM, Kimi — which, fairly or not, raised sourcing and governance questions for some regulated buyers regardless of the license text. A US-origin, Apache-2.0, frontier-scale model removes one of those variables from the decision, and for teams in regulated or public-sector settings that variable was often the one doing the blocking.
The other genuinely practical feature is controllable thinking effort. Developers set a reasoning budget per request — on a scale from roughly 0.2 to 0.99 — trading compute and latency for depth on a call-by-call basis. That moves a lever which has usually lived at the model level (pick the big model or the small one) down to the individual request. For agentic and RAG workloads, where most steps are cheap and only a few are hard, spending reasoning only where it pays is a real cost control — and it is exactly the kind of knob you rarely get to turn on a hosted API you do not own.
The honest caveat is scale. The full-precision checkpoint wants something like 2TB of aggregate GPU memory — on the order of eight B300s or sixteen H200s. An NVFP4 quant drops that toward 600GB but needs Blackwell-generation hardware to run at all, so existing H100 clusters are out, and even aggressive community quants land around 270GB. "Open weights" is not "runs on your workstation." Owning the model is a licensing and provenance win; running it is still an infrastructure project, and the teams who actually benefit are the ones who already have — or will rent — serious GPU capacity, not everyone who can click download.
The signal is that the open-weight frontier is no longer a one-country story, and that "open" is finally starting to mean the license as much as the download. If provenance or licensing was the thing keeping a self-hosted model off your shortlist, that objection just got weaker — which is reason enough to re-open the question rather than assume last quarter's answer still holds.
The move is a concrete test, not a rewrite of your stack. Validate the reasoning-budget behavior and the real per-request inference cost on your own workload before committing, because a 975B model you host is a bill you own end to end — capacity, quantization, latency, and all. The weights are free. Everything downstream of them is still a decision.
If a private or on-prem model has been stuck behind licensing, provenance, or cost questions, this is a good moment to re-run the numbers on your own workload rather than the vendor's benchmark. If your team is trying to move an AI use case from demo to deployment, METECH helps scope, build, and validate the first working system in 2-3 weeks.