Mira Murati's Thinking Machines Shipped Its First Model. It's a 975B Open Weight Aimed at China's Open Labs.

TL;DR

Thinking Machines Lab, the startup from former OpenAI CTO Mira Murati, finally shipped a model. It is Inkling, and the headline is not the size, it is the license. This is an Apache 2.0 open-weight Mixture-of-Experts with 975B total parameters, only 41B of them active per token, pretrained on 45 trillion tokens of text, images, audio, and video. It reasons natively over text, image, and audio, stretches to a 1M-token context, and ships weights you can download and fine-tune today. The company says plainly it is "not the strongest overall model available today, open or closed." That is the point, and it is aimed squarely at the Chinese open-weight labs. Here is what it does, what it does not, and whether you can run it.

What Murati actually shipped

The spec sheet reads like a statement of intent. Inkling is a decoder-only multimodal Mixture-of-Experts with 975 billion total parameters and 41 billion active. Under the hood it runs 256 routed experts plus 2 shared experts, and routes each token to the top 6 of those 256, with the 2 shared experts firing on every token. It was pretrained on 45 trillion tokens spanning text, images, audio, and video, accepts text, image, and audio as input, emits text, and supports a context window up to 1 million tokens (Tinker exposes 64K and 256K options today). There is also an Inkling-Small preview: a 276B-parameter MoE with just 12B active, for people who want the same recipe at a lighter weight.

That 41B-active figure is the whole trick, and it is worth slowing down on. In a Mixture-of-Experts a small router decides, per token, which handful of expert sub-networks to actually run, so you store all 975B parameters but only spend compute on a fraction. Picture a library with 256 specialist reference desks: every question is routed to the 6 desks most likely to know the answer, plus 2 general-desk librarians who field everything, so you keep an enormous library on the payroll but only ever wake a handful of desks per question.

The payoff is cost. Thinking Machines says Inkling reaches equivalent coding performance to NVIDIA's Nemotron 3 Ultra while burning about a third of the tokens, and it is launching the hosted version at a 50% discount for a limited time.

The bet: a model you shape yourself

A US lab open-sourcing a near-trillion-parameter model is the part that should make you sit up. Murati's old employer keeps its frontier weights locked, as do Anthropic and Google. Her first public move as a founder is to give the weights away. Call it a philosophy or the most expensive subtweet in AI history, either way it is a real strategic wager.

The wager is that customization beats the one-size-fits-all API. Thinking Machines argues that AI trained centrally by one company and then set in stone underperforms a model an organization can shape for itself, and the business model follows directly: Tinker, the company's fine-tuning and hosting platform, is where the revenue is meant to come from. Open weights are the on-ramp. Inkling is available for fine-tuning on Tinker on day one, and is also served through Together AI, Fireworks, Modal, Databricks, and Baseten. Give the model away, then sell the workshop where people bend it to their own data.

How good is it, honestly

Here is the refreshing part: the launch post does not pretend Inkling tops the charts. It leads where it leads and trails where it trails, and says so out loud. On Thinking Machines' own numbers, run at maximum thinking effort of 0.99, it is strong on reasoning and multimodal understanding: 97.1% on AIME 2026 math, 91.4% on VoiceBench, 87.2% on GPQA Diamond, and 77.6% on SWE-bench Verified.

The soft spots are just as clear. On SimpleQA Verified, a factual-recall test, Inkling scores 43.9%, well under the 57.0% reported for DeepSeek V4 Pro. On Terminal Bench 2.1, an agentic-coding harness, it trails GLM-5.2 and Kimi K2.6 by a wide margin. Where it does lead the open field is robustness: it posts the best FORTRESS adversarial score among open-weight models at 78.0%, and a positive AA Omniscience (hallucination-resistance) score while Nemotron 3 Ultra, Kimi K2.5, and DeepSeek V4 Pro all land in negative territory. As always, these are vendor-run numbers on a vendor harness, so the only score that pays your bills is the one you measure on your own workload. We put GLM-5.2, the model Inkling most directly challenges, through its paces in our earlier writeup.

Can you actually run it

Now the homelab reality check, because open weights do not mean small. The full BF16 checkpoint wants roughly 2TB of VRAM, which is a multi-node datacenter answer, not a desk one. The NVFP4-quantized build drops that to about 600GB, still a small cluster. This is the same trap we flagged with Tencent's Hy3: MoE saves you compute per token, but you still have to hold every parameter in memory, because the router might reach for any expert at any moment.

The community is already chipping away at it. Unsloth has dynamic 1-bit GGUF quantizations that, by its own reporting, cut VRAM by around 95% while retaining roughly 74.2% of the model's top-1% accuracy. That is the difference between "rent a cluster" and "maybe, on a serious multi-GPU rig, if you can stomach the quality hit." For most people the realistic path today is the hosted endpoints or a Tinker fine-tune, not weights on your own metal, at least until the quant work matures.

Key Takeaways

• The debut: Thinking Machines Lab, Mira Murati's startup, shipped its first model, Inkling, on July 15, 2026, and released it as Apache 2.0 open weights.
• The architecture: a 975B-parameter multimodal Mixture-of-Experts with 41B active, 256 routed experts (top-6) plus 2 shared, trained on 45T tokens of text, image, audio, and video, with up to a 1M-token context.
• The bet: that a model you fine-tune yourself beats a locked frontier API. Tinker, the company's customization and hosting platform, is the business behind the free weights.
• The honesty: Thinking Machines says outright it is "not the strongest overall model." It leads open models on safety and factual robustness (78.0% FORTRESS) but trails on factual recall (43.9% SimpleQA) and agentic coding (Terminal Bench).
• Running it is hard: BF16 needs ~2TB VRAM and NVFP4 ~600GB; community 1-bit quants cut VRAM ~95% at a real accuracy cost. Hosted endpoints and Tinker are the practical path for now.
• The bigger signal: a US lab open-sourcing a near-1T model, aimed straight at the Chinese open-weight leaders, from the person who used to run OpenAI's technology.

Sources: Thinking Machines Lab: Introducing Inkling, Inkling model card (Hugging Face), Welcome Inkling (Hugging Face blog), TechCrunch, MarkTechPost, OfficeChai (benchmarks)