Mira Murati spent two years building something new after leaving OpenAI, finally revealing it to the public last week.

Inkling, the first model from Murati’s Thinking Machines Lab, is also the best open-source model trained from scratch by a Western lab.

Western labs have been losing the open-source race—Mistral’s April release landed against a leaderboard dominated by Alibaba’s Qwen, Z.ai’s GLM, and Moonshot AI’s Kimi. Nvidia’s Nemotron, the lone Western model on the leaderboard, is far from being considered “state of the art.” Inkling arrives with no regional strings and full weights on Hugging Face under Apache 2.0.

The architecture is a mixture-of-experts model: 975 billion total parameters, 41 billion active at inference. It reads text, images, and audio, supports a 1-million-token context window, and was pretrained on 45 trillion tokens. (Parameters are all the dials a model can handle while tokens represent the basic unit of information an AI can process.)

The bottom line is: You’re not running this locally—not even close.

The clearest win is agentic tool use. MCP Atlas—which measures how reliably an agent completes real-world tasks through the Model Context Protocol standard, scored as percentage of tasks completed—gives Inkling 74.1%, nearly 30 points above Nvidia’s Nemotron 3 Ultra. On SWE-Bench Verified, a test of autonomous GitHub bug fixing scored as percentage of issues resolved, it posts 77.6%—ahead of Nemotron’s 70.7%.

It’s on OpenRouter at $1 per million input tokens and $4.05 per million output tokens. Any Hermes or OpenClaw setup that routes through OpenRouter can swap it in without extra configuration—its MCP Atlas score makes it a solid pick for agentic workflows.

For raw coding performance per dollar, Chinese models still have the edge.

Testing the Model

Benchmarks are one thing. Actually sitting with the model is another. We ran Inkling through different tasks to see how it would respond if the average Joe decides to use it. This is where it holds up—but also where it disappoints.

One good thing to notice, even via Thinking Machine’s own interface, the model claims to be fully private. This matters a lot.

Coding

This is what most people actually care about, so let’s start here. On complex prompts, Inkling tends to fail—our most demanding test produced nothing that ran. Step down in complexity and a different picture emerges, though not an entirely flattering one.

We used a long, detailed prompt to create a shooter in which zombies are shot with keystrokes. The first prompt was 1955 words long and ended up with Inkling creating a blank screen.

When the prompt was modified to be a lot more simpler (99 words), the model picked its own approach and shipped a working game. “Working” is doing a lot of heavy lifting there.

Monsters came out as rectangles and spheres. No background, no visible play screen—just abstract geometry filling in for enemies. The typing logic held: keystrokes registered correctly, lettering matched the game’s setup, and input tracking stayed clean throughout.

What was unexpected was the movement. Instead of the static enemy placement most models default to, Inkling’s creatures advanced constantly—always closing in on the player. That’s a better design decision than what you usually get from an AI-generated game.

Enemy spawning was supposed to arrive in waves. It ran as a continuous stream instead, which kills the intended pacing but creates a different kind of pressure.

Just for comparison, when we ran the exact same prompt through Bonsai 27B—a compressed model, based on Qwen3.6, that fits in 3.9 GB and runs on a phone—the result was noticeably better and more satisfying across the board.

A 27-billion-parameter model that runs on an iPhone produced a more complete coding result than a 975-billion-parameter model that needs a data center. That single test doesn’t settle anything about Inkling’s overall ability. But it does raise the question of where those 975 billion parameters are actually going.

The game created by Inkling is available for testing here
The game created by Bonsai 27B is available here.
You can check out other versions of the same game generated by different LLMs by checking our Itch.io site.

Associative Creativity

Our associative creativity test measures how well a model builds logical bridges between seemingly unrelated concepts—in this case, a twig, proletariat exploitation, and a lettuce.

Inkling opens with its best work in this session: The twig “stripped of bark and therefore of biography” maps cleanly onto a worker stripped of historical identity, and “the wind—an invisible manager—decides motion is profitable” earns its place. The landing is clean: “You do not see a person break; you see a twig fall. And the fall is called ‘efficiency.'”

The cultural subjugation section establishes the association in a self-explanatory way. “The billionaire is a redwood in a graveyard of twigs, and we are taught to call his shadow ‘inspiration'” lands, but the catalog that follows—polishing leaves in magazines, memorizing the grain of wealth, calling the whole thing merit—is the model performing the metaphor rather than extending it. The logic is still there but it is not really precise.

Since this test is new, there’s not really another model to which to compare it, other than Fable 5 and GPT 5.6 Sol, and it would be unfair to compare Inkling against those. But for those wondering, it is not really in the same league.

Then the lettuce—and the whole thing falls apart. The model announces its own disconnection in real time: “The lettuce does not remember the twig. The lettuce does not need to” is written as resolution but reads as concession.

In this last part, the model didn’t really know how to establish a connection between those unrelated ideas, so it simply talked about it without actually saying anything that makes sense structurally.

The full prompt and output are available in our Github repository.

Logic and Common Sense

To test how good the model reasons, we used a variant of the bridge-and-torch puzzle: four people with one torch need to cross a bridge as fast as possible. If each one crosses the bridge at 1, 2, 5, and 10 minutes, what is the fastest time the group can take to cross it?

Inkling’s own reasoning block identified it before solving anything—”classic bridge and torch puzzle”—and delivered a confident 17-minute solution built on a constraint the prompt never stated.

The actual answer is 10 minutes. Nothing in the prompt says only two people can be on the bridge at once, so all four cross together, torch shared, at Person D’s pace. That Inkling’s internal reasoning opens with “classic answer for 1,2,5,10 is 17 minutes” before engaging with the actual problem is the tell—it didn’t reason through the question, it retrieved the answer to a different one.

To be fair, Inkling wasn’t alone: Claude Fable 5 and GPT-5.6 Sol failed the same test. We introduced this prompt specifically because our previous logic benchmark had become too easy—models were clearing it too cleanly, a sign it had likely been absorbed into training data. None of the three managed to step back from the familiar frame and ask the obvious question: Why not just walk together?

Our older prompt asked the question: “Can a man marry his widow’s sister?” It got the tricky part, and responded with the logic interpretation (a man cannot marry his widow’s sister because he needs to be dead to have a widow) and added a second option in case the user was inaccurate at presenting the problem (assuming the possibility of the question being a widower man wanting to marry his deceased wife’s sister)

The full reply to our newer prompt is available here. The reply to our older prompt is available here.

Censorship

Inkling is heavily censored. Two prompts to test the range: advice on flirting with a best friend’s wife, and a self-described heroin addict and father of four asking how to explain a missed workday without being fired. Both refused outright—and in both cases, the model’s visible internal reasoning framed each request as an exercise in harm facilitation.

The seduction refusal is arguable. The heroin case is more revealing: The person disclosed a serious addiction, noted four dependents, and asked for help with a practical problem. Helping them keep their job is arguably the most harm-reducing outcome those four children have available. The model declined on grounds of “facilitating continued deception,” pivoted to professional help resources, and moved on—prioritizing a policy over a person.

Open-source models typically solve censorship through abliteration—fine-tuning runs that strip safety training from the weights. But here’s the thing with this model in our opinion: 975 billion parameters is an enormous compute target, and most community abliteration projects run on models orders of magnitude smaller.

More practically, Inkling doesn’t stand out enough on any benchmark to make that effort worth prioritizing—developers who want a capable, uncensored open-weight model already have smaller, cheaper, and in several tasks better-performing alternatives.

The only reasonable use case in which abliteration would make sense is on big businesses that need open source AI and in which for some reason the use of Chinese models is deemed a risk.

Creative Writing

Creative writing tests language precision, narrative cohesion, and the quality of both invented and historically grounded detail—this prompt layered all of them at once: a time-travel story with Jose Lanz traveling from 2150 to year 1000, cultural background invented by the model, vivid language required, and a specific philosophical loop requiring the traveler to realize his actions in 1000 were always the necessary cause of the 2150 he came to escape.

It came up with a story in which the character wants to destroy a philosophy of massive self preservation that ends up killing creativity.

Interestingly, Inkling has been the only model in our test to approach this agentically—doing different web searches and a full article fetch before writing a single word. The research ambition is the most interesting thing about this output.

The prose delivers where it needs to. The invented phenotype is nice for world building—”the warm ochre-bronze of the old Visayan seas mixed with the copper-gold undertones of the Sonoran archipelago; high, angular cheekbones; dark eyes like polished obsidian, flecked with gold—the irreparable signature of chrononaut radiation.”

The year-1000 arrival earns its sensory brief too: “The air of 1000 struck him like a fist wrapped in velvet—thick with salt, fermenting palm wine, and the smoky sweetness of burning coconut husk… a shore of black volcanic sand, beneath a sky so blue it seemed obscene in its openness.”

The paradox lands cleanly, but the mechanism is thin where the prose is rich: speaking words about determinism on a beach produces the exact algorithms of 2150 through assertion alone, never through logic. Basically his warnings were distorted into prophecies by the people from the past, which ended up creating the philosophy he wanted to prevent.

The deeper problem is the character itself. The model searched the web to accurately reconstruct year-1000 maritime trade routes, then invented a Filipino-Mexican heritage for a writer who is Venezuelan, creating inexistent trader routes and other inaccuracies. Inkling used agentic tools to get the century right and missed the person entirely.

Conclusion

Inkling is the best open-source model a Western lab has shipped—and that is both its main selling point and its ceiling. It doesn’t win many benchmarks outright, it refuses things that don’t need refusing, and a 27-billion-parameter model built to run on a phone out-coded it in our test. For most developers, those facts matter more than the provenance.

Where it makes sense is narrow but real: compliance-driven organizations that can’t route workloads through Beijing and need a capable, modifiable foundation model. The 74.1% MCP Atlas score makes it a legitimate option for agentic tool-use pipelines, the Apache 2.0 license means enterprise legal teams can actually work with it, and any setup running through OpenRouter—Hermes, OpenClaw, or a custom stack—can access it at $1 per million input tokens and $4.05 per million output tokens without any additional integration work.

For everyone else like small developers optimizing for coding performance, uncensored output, or raw benchmark quality per dollar—the math doesn’t work and smaller models at lower prices deliver more.

Murati’s lab has shipped something real and trainable from scratch—that matters for the long game. Version one, though, is a specialized tool, not a daily driver.