Mira Murati's New AI: Why Inkling Isn't Trying to Be

[00:00:00]A brand new lab shipped a 975 billion parameter model.

[00:00:04]Then admitted it isn't the strongest.

[00:00:07]That concession is the entire strategy.

[00:00:10]This is the most honest artificial intelligence launch of the year.

[00:00:14]The lab is Thinking Machines, founded by Mira Murati, the former Chief Technology Officer at OpenAI.

[00:00:20]On July 15th, it shipped its first foundation model, Inkling, and put the full weights in the open.

[00:00:26]975 billion parameters. Anyone can download.

[00:00:31]Two questions hang over it, and neither answer is the benchmark.

[00:00:35]Why open by admitting it isn't the best?

[00:00:37]And if the weights are open, why can't you run them?

[00:00:41]Underneath Inkling is a mixture of experts.

[00:00:44]A giant model that only uses a slice of itself at a time.

[00:00:47]Picture 256 specialists.

[00:00:50]A word comes in, a router wakes just six plus two that are always on.

[00:00:55]Everyone else stays dark.

[00:00:57]On paper, 975 billion parameters.

[00:01:01]Doing the work on any given token, only 41 billion.

[00:01:05]A 975 billion parameter library with a 41 billion working brain.

[00:01:11]That's how a model this size runs at the same speed. You never light up all 256 at once.

[00:01:19]And it scans a full million tokens in one pass.

[00:01:22]A whole code base or a long recording in a single shot.

[00:01:25]Inkling doesn't stop at text.

[00:01:28]It trained on 45 trillion tokens of text, images, and audio.

[00:01:32]All three go straight into the same reasoning core.

[00:01:35]Most models bolt vision or audio on afterward as a separate adapter.

[00:01:40]Inkling takes image patches and raw audio natively.

[00:01:43]Reasons over them together.

[00:01:45]Answers in text.

[00:01:47]Hand it a screenshot, a chart, or a voice clip.

[00:01:50]It treats them as one problem, not three.

[00:01:53]For a small team, that's one model to run and one bill to reason about.

[00:01:58]No stitching a vision service and a speech service around a language model.

[00:02:02]But reasoning has a price.

[00:02:04]Every thinking model burns extra tokens to work a problem through and you pay for everyone.

[00:02:09]Inkling puts that on a dial.

[00:02:12]You pick the effort from none up to max.

[00:02:14]Seven settings.

[00:02:16]A quick look up stays cheap.

[00:02:18]A hard proof gets the deep pass.

[00:02:20]And the efficiency number is the real headline.

[00:02:23]On terminal bench thinking machine says Inkling matches Nemotron 3 Ultra using about a third of the tokens.

[00:02:30]For a small team watching its inference bill, fewer tokens at the same quality beats topping a leaderboard.

[00:02:37]Here's the catch, the headline skip.

[00:02:39]You cannot run this on your own machine.

[00:02:42]Open weights does not mean download and run.

[00:02:44]The file is the easy part.

[00:02:46]To load Inkling at full precision, you need about 2 terabytes of VRAM.

[00:02:51]Eight of Nvidia's B300 chips or 16 H200s.

[00:02:56]Even the compressed NVFP4 version wants around 600 GB.

[00:03:01]Your laptop has maybe 8 to 96 GB total.

[00:03:05]So it runs on data center Blackwell and Hopper GPUs, the B300s and H200s.

[00:03:11]Not your laptop and not a single gaming card.

[00:03:14]Open weights got you the keys.

[00:03:16]They did not get you the forklift.

[00:03:18]So how do you actually use it?

[00:03:21]You do, just not the way you assumed.

[00:03:23]The inference providers you already reach for, Together AI, Fireworks, Model, Databricks and Base 10 are hosting Inkling from day one.

[00:03:32]The open weights become a normal API call on the stack you already pay for.

[00:03:37]Cheaper, faster, open model inference with no GPU purchase.

[00:03:41]Want to shape it to your own domain?

[00:03:44]Thinking Machines runs a managed fine-tuning service Tinker that trains on their hardware, not yours.

[00:03:50]They launched it at half price with a free playground. And if 975 billion is too heavy, there's Inklings more.

[00:03:58]A 276 billion parameter preview. 12 billion active tuned for coding.

[00:04:05]Treat the big model as a blank base you carve down to your own problem.

[00:04:09]Quick word about Hostinger.

[00:04:11]Four ways to run AI agents on your own infrastructure.

[00:04:15]AI agents, connector, open claw, and Hermes agent. Use code DIY smart code for an extra 10% off at hostinger.com.

[00:04:24]Okay, back to the video.

[00:04:27]Now the benchmarks.

[00:04:29]Inkling isn't on top, and that's on purpose. Against eight other frontier models, it lands mid-pack almost everywhere.

[00:04:37]On the AIME math test, 97.1%.

[00:04:41]GLM 5.2 hits 99.2.

[00:04:44]GPT 5.6 still 99.9.

[00:04:48]Once we bench 77.6.

[00:04:51]Behind Kimmy K 2.6 at 80.2.

[00:04:54]Well behind Claude Fable 5 at 95.

[00:04:57]So it loses to the open models, too.

[00:05:00]Why cover it?

[00:05:01]Here's the reframe.

[00:05:03]Thinking Machines told you up front it isn't the strongest.

[00:05:06]The score was never the pitch. The pitch is an Apache licensed base you can host, fine-tune, and reshape. A different product than a number on a chart.

[00:05:16]Which is why the license beats the leaderboard here.

[00:05:19]Plenty of models call themselves open, then bury a custom license that blocks commercial use or gates how you deploy.

[00:05:26]Inkling ships under plain Apache 2.0, confirmed right on the model card.

[00:05:31]Build on it, ship a product, no lawyer on the fine print.

[00:05:35]And that matches the honesty from the very first line.

[00:05:38]A lab that tells you where it loses is easier to trust on where it wins, like native audio and controllable cost. The honesty is the whole product.

[00:05:47]Real quick, the dynamist.ai community is where I go deeper on this stuff.

[00:05:52]So, here's the question.

[00:05:54]Inkling openly admits it's not the strongest model.

[00:05:57]Is that honest positioning a smart strategy or a cop-out?

[00:06:01]Tell me below and subscribe for more open model news.