Summary
- Inkling, a 975-billion-parameter open-source AI model, was launched by the Thinking Machines Lab on July 15, 2024.
- This marks the first significant release from Mira Murati's lab since her departure from OpenAI in September 2024.
- Available on OpenRouter, the model costs $1 per million input tokens and $4.05 per million output tokens, though it faces competition from models that offer better performance at similar or lower prices.
After two years of development post-OpenAI, Mira Murati unveiled Inkling to the public last week.
This model, the inaugural release from Murati's Thinking Machines Lab, is touted as the leading open-source model developed from scratch by a lab in the West.
Western developers have struggled in the open-source sector, with Mistral's recent release facing off against strong competitors like Alibaba’s Qwen, Z.ai’s GLM, and Moonshot AI’s Kimi. Nvidia’s Nemotron, the only Western model on the leaderboard, is not considered cutting-edge. Inkling is released without regional restrictions and is available on Hugging Face under the Apache 2.0 license.
Its architecture features a mixture-of-experts design with a total of 975 billion parameters, of which 41 billion are active during inference. The model can process text, images, and audio, supports a context window of up to one million tokens, and was pretrained using 45 trillion tokens. (Parameters are the variables a model can manipulate, while tokens are the basic units of information it processes.)
However, it’s important to note that running this model locally is not feasible.
One of Inkling’s standout features is its ability to perform agentic tasks. According to the Model Context Protocol (MCP) Atlas, which evaluates the reliability of an agent in completing real-world tasks, Inkling scored 74.1%, significantly outperforming Nvidia's Nemotron 3 Ultra by nearly 30 points. On SWE-Bench Verified, a benchmark for autonomous bug fixing on GitHub, Inkling achieved a score of 77.6%, again ahead of Nemotron's 70.7%.
Source: Thinking MachinesInkling is accessible on OpenRouter for $1 per million input tokens and $4.05 per million output tokens. It can be seamlessly integrated into any Hermes or OpenClaw setup using OpenRouter without additional configuration due to its MCP Atlas score, making it a strong candidate for agentic workflows.
However, Chinese models still hold an advantage when it comes to raw coding performance per dollar spent.
Evaluating the Model
While benchmarks provide insights, hands-on experience with the model reveals both strengths and weaknesses. We tested Inkling with various tasks to gauge how it would perform for the average user, yielding mixed results.
Notably, Inkling claims full privacy even through Thinking Machines’ interface, which is a significant consideration.
Coding
Coding performance tends to be a primary concern for many users, so we began our evaluation here. Inkling struggled with complex prompts—our most challenging test yielded no functional output. However, simplifying the prompt produced better results, though not without shortcomings.
Initially, we tasked Inkling with creating a shooting game involving zombies with a 1955-word prompt, which resulted in a blank screen. When we shortened the prompt to 99 words, the model generated a playable game, albeit with limitations.
The game featured enemies represented as basic shapes, lacking a proper background or play area—merely abstract geometry stood in for the zombies. Despite these flaws, the game's input tracking functioned correctly, allowing keystrokes to register as intended.
Interestingly, the enemy characters in this version moved continuously toward the player, a design choice that was more engaging than the static placements typical of AI-generated games.
However, the spawning of enemies was intended to occur in waves but instead happened continuously, disrupting the desired pacing and creating a different type of challenge.
In a comparative test, when we used the same prompt with Bonsai 27B, a smaller model based on Qwen3.6 that operates on mobile devices, the output was noticeably superior and more satisfying overall.
The Bonsai 27B, a 27-billion-parameter model that can run on an iPhone, produced a more complete coding result than Inkling, which requires a data center. While this single test does not definitively determine Inkling's overall capabilities, it raises questions about the use of its 975 billion parameters.
The game developed by Inkling is available for testing here.
The game created by Bonsai 27B can be found here.
Additional versions of the same game generated by various LLMs are accessible on our Itch.io site.
Associative Creativity
In assessing associative creativity, we examined how well the model connected seemingly unrelated concepts: a twig, proletariat exploitation, and lettuce.
Inkling performed best with the twig, articulating a connection to a worker stripped of identity, and effectively conveyed the metaphor of the wind as an invisible manager determining profitable movement. The conclusion was striking: "You do not see a person break; you see a twig fall. And the fall is called 'efficiency.'"
However, while the model established associations, it struggled to maintain precision in its catalog of ideas, veering into metaphor without extending logical connections.
This test is novel, and while it can be compared to Fable 5 and GPT 5.6 Sol, it would be unfair to directly pit Inkling against those models, as they belong to a different caliber.
The connection with lettuce ultimately faltered, with the model admitting, "The lettuce does not remember the twig. The lettuce does not need to," which reads more as a concession than a resolution.
The complete prompt and response from this test are accessible in our Github repository.
Logic and Common Sense
To evaluate reasoning capabilities, we employed a variation of the bridge-and-torch puzzle: four individuals with a single torch must cross a bridge in the shortest time possible, taking 1, 2, 5, and 10 minutes respectively. What’s the quickest time for the group to cross?
Inkling recognized the puzzle immediately, labeling it a "classic bridge and torch puzzle" and confidently providing a 17-minute solution based on an unstated constraint. However, the correct answer is 10 minutes, as nothing in the prompt limits the number of people crossing the bridge simultaneously.
Inkling's initial reasoning, which suggested a 17-minute outcome, indicates it retrieved an answer for a different scenario rather than engaging with the problem at hand. It is worth noting that neither Claude Fable 5 nor GPT-5.6 Sol succeeded in this test either, as they failed to consider the possibility of all four individuals crossing together.
Our previous logic prompt, which asked: “Can a man marry his widow’s sister?” was effectively interpreted by Inkling, which provided a logical explanation about the impossibility of a man marrying his widow’s sister, adding a second interpretation for clarity.
The complete responses to our current and previous prompts can be found here and here, respectively.
Censorship
Inkling exhibits significant censorship. In testing its boundaries, we posed two prompts: one seeking advice on flirting with a best friend's wife, and another from a self-identified heroin addict with dependents asking how to explain a missed workday. Both prompts were rejected outright, with the model framing each request as harmful facilitation.
The refusal to provide flirting advice is debatable, but the denial in the case of the heroin addict is more revealing. The model declined to assist, prioritizing policy over the individual’s urgent needs, even as the request came from someone in a vulnerable situation.
Typically, open-source models address censorship through a process called abliteration, which removes safety training from the model weights. However, given Inkling’s scale of 975 billion parameters, most community efforts in this area focus on models much smaller.
From a practical standpoint, Inkling does not distinguish itself in benchmarks enough to warrant such efforts—developers seeking robust, uncensored open-weight models already have access to smaller, more cost-effective alternatives that perform better in several tasks.
The only plausible scenario for abliteration would involve large organizations requiring open-source AI while avoiding the use of Chinese models for various reasons.
Creative Writing
We assessed creative writing through a prompt requiring narrative coherence, precision in language, and historical accuracy, asking for a time-travel story featuring a character named Jose Lanz traveling from 2150 back to 1000 AD. The narrative demanded vivid descriptions and a philosophical twist where Lanz realizes his actions in 1000 were the cause of the 2150 he aimed to escape.
Inkling uniquely approached this task by conducting web searches and retrieving articles before crafting its narrative, showcasing its research capabilities.
The prose was compelling, with striking imagery such as the character’s description: "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."
While the paradox was effectively conveyed, the underlying logic was weak, leading to a narrative where the character’s warnings became distorted prophecies rather than coherent arguments. Additionally, while the model accurately reconstructed historical maritime trade routes, it inaccurately assigned a Filipino-Mexican heritage to a Venezuelan character, leading to fictionalized trade routes.
Conclusion
Inkling stands as the leading open-source model produced by a Western lab, representing both its primary advantage and limitation. While it doesn’t dominate many benchmarks, it does excel in specific areas. However, a smaller 27-billion-parameter model outperformed it in our coding tests, raising questions about its practical application for most developers.
Its best use scenarios are likely within compliance-driven organizations that require a modifiable model without risking reliance on Chinese technology. The 74.1% MCP Atlas score indicates it can be effectively utilized in agentic tool-use scenarios, and its Apache 2.0 license allows for legal use in enterprises. Accessible through OpenRouter at a reasonable cost, it integrates easily into existing setups.
However, for smaller developers focused on coding efficiency, uncensored outputs, or superior benchmark performance, Inkling may not be the best fit, as smaller and cheaper models often deliver better results.
Overall, Murati's lab has made a meaningful contribution by launching a trainable model from scratch, which is significant for future developments, but the current version is more of a specialized tool than a general-use solution.