Mira Murati's Thinking Machines Just Shipped Inkling, a 975B Open-Weights Model Under Apache 2.0
Quick verdict
Thinking Machines Lab, the startup Mira Murati started after leaving OpenAI, just released its first open-weights model. Inkling is a 975B-parameter mixture-of-experts model with 41B active per token, it takes text, images, and audio, and it ships under Apache 2.0 with a 1M-token context. On the Artificial Analysis Intelligence Index it scores 41, a hair above Nvidia's Nemotron 3 Ultra at 38, but it does not top the closed frontier and the lab is not pretending it does. The pitch is a customizable base you can fine-tune and run yourself, not a leaderboard trophy. For a company sitting on this much money to put its flagship out under one of the most permissive licenses around, that is the actual news.
What actually shipped
Inkling is a sparse MoE model, 975B total parameters with about 41B active per token, trained on 45 to 48 trillion tokens. It handles text, image, and audio input and returns text. The open weights carry a 1M-token context window, while the hosted Tinker API caps it at 256K. There is also a smaller sibling, Inkling-Small, at 276B total and 12B active, for people who want something lighter. Everything is Apache 2.0.
The architecture has a few choices worth flagging. It uses hybrid sliding-window attention at a 5:1 local-to-global ratio and relative positional encoding instead of RoPE, which a few people reading the release flagged as a possible hint of where architectures go next. It leans on DeepSeek-style auxiliary-loss-free load balancing, two shared experts, unusually large short-convolution layers, and 8 multi-token-prediction heads for speculative decoding. Mira Murati posted the launch, and John Schulman and Soumith Chintala both weighed in on the engineering.
On benchmarks the picture is solid rather than dominant. The launch thread and third-party trackers put it at 41 on the Artificial Analysis Intelligence Index, 1238 Elo on GDPval-AA v2, and #9 overall at 1257 Elo on Design Arena's agentic web-app board. On τ³-Banking, an agentic tool-use test, it lands at 24%, which is modest. So this is a capable general model that reads as a foundation to build on, not a coding or agent specialist you drop in and win with.
Day-zero support was the real flex
The thing that stood out was how much of the serving stack was ready on launch day. vLLM reported 380 tokens per second per user on 4x GB200 with multi-token prediction, shipping NVFP4 and BF16 checkpoints. SGLang added native support for the architecture, including the short-conv layers and an MXFP8 KV cache. Modal built a custom speculator it says gives 67% higher throughput. Baseten, Databricks, and Hugging Face all announced support the same day, and Unsloth put out 1-bit GGUF quants that compress the model 86%, from roughly 1.9TB down to 270GB.
That last number is the reality check. Even the aggressively quantized build is 270GB, and the full weights are measured in terabytes. Inkling is open, but open is not the same as local. Almost nobody is running the 975B version on hardware they own. For most people the practical path is a hosted provider, exactly like it was with GLM, Kimi, and the other big open releases this year.
Why it matters
Two things make this land. First, the source. Thinking Machines raised one of the largest seed rounds in history and stayed quiet for a long stretch, so a lot of people expected a closed API product aimed straight at OpenAI and Anthropic. Instead the debut is Apache-2.0 open weights. That license lets you use Inkling commercially, fine-tune it, and ship it inside a product without the usage clauses some other open releases attach.
Second, the positioning. The lab is explicit that Inkling is a customizable multimodal base, not a benchmark-maxed flagship. That is a different bet than the one most frontier labs make. Rather than chase the top of the intelligence index, it hands you a large, permissively licensed model and a fine-tuning API and lets you shape it. For teams that care about controlling their own model instead of renting whatever a vendor exposes, that is a more useful starting point than another closed endpoint, even if the raw scores trail the leaders.
If you are weighing the broader open-versus-closed question, or just trying to keep your model spend sane while these releases keep landing, our take on the best open-source AI models in 2026 and on routing requests to the cheapest capable model both cover the practical side.
Video: Thinking Machines and the open-weights bet
Context on Mira Murati's lab and where an Apache-licensed frontier-scale model fits.
FAQ
Can I run Inkling on my own machine?
Not the 975B version, realistically. Even the 1-bit GGUF quant is about 270GB, and the full weights run into terabytes, so most people will use it through a hosted provider like Baseten, Databricks, or a vLLM/SGLang endpoint. The smaller Inkling-Small at 276B is lighter but still well beyond a normal laptop. Our roundup of open-source AI models covers what is actually runnable at home.
Is Inkling better than GPT-5.6 or Claude?
Not on the headline benchmarks. It scores 41 on the Artificial Analysis Intelligence Index, which is strong for an open model but below the closed frontier, and its agentic tool-use scores are modest. The case for Inkling is the Apache-2.0 license and a fine-tunable multimodal base, not beating the closed leaders head to head.
What does the Apache-2.0 license let me do?
You can use it commercially, modify it, fine-tune it, and ship it inside your own product without the extra usage restrictions some open releases carry. That permissiveness, from a lab this well funded, is a big part of why the release got attention beyond the raw scores.
Sources
- Thinking Machines Lab - Inkling launch thread, specs and benchmarks
- Mira Murati - announcing the Inkling release
- John Schulman - on Inkling's training and design
- Soumith Chintala - on the serving stack and speculative decoding
Further reading
Try all the models mentioned in this article
Admix gives you GPT-5, Claude, Gemini, and 350+ AI models in one app. Compare responses side by side. Free to start.
Start free on Admix