After two years of dedicated development following her departure from OpenAI, Mira Murati, a pivotal figure in the artificial intelligence landscape, has finally revealed her new venture, Thinking Machines Lab, and its inaugural large language model, Inkling. Unveiled to the public last week, Inkling enters the competitive arena with a significant claim: it is positioned as the foremost open-source model trained from scratch by a Western laboratory, challenging the dominance of Eastern counterparts that have increasingly captured the lead in this crucial sector of AI innovation.
Inkling’s arrival is particularly noteworthy given the current global dynamics of open-source AI development. Western laboratories have recently found themselves lagging in the race, with many of the top-performing open-source models originating from Asia. For instance, Mistral’s highly anticipated April release found itself vying for attention against a leaderboard heavily populated by models such as Alibaba’s Qwen, Z.ai’s GLM, and Moonshot AI’s Kimi. Nvidia’s Nemotron, though a prominent Western entry, has struggled to be considered "state of the art" in comparison. Inkling, however, aims to disrupt this trend, offering full weights on Hugging Face under the permissive Apache 2.0 license, free from regional constraints often associated with models from specific geopolitical blocs.
The Genesis of Thinking Machines Lab and Inkling
Mira Murati’s journey from her influential role as Chief Technology Officer at OpenAI to founding Thinking Machines Lab provides critical context for Inkling’s debut. At OpenAI, Murati was instrumental in the development and scaling of groundbreaking models like ChatGPT and DALL-E, playing a key role in bringing advanced AI to a mainstream audience. Her decision to embark on an independent venture underscored a belief in a different path for AI development, potentially one that prioritized open-source principles or novel architectural approaches. The two years of quiet development suggest a meticulous, deliberate process aimed at building a robust foundation model designed to stand apart.
The strategic importance of open-source AI cannot be overstated. It fosters transparency, democratizes access to powerful technology, and accelerates innovation by allowing a global community of developers to inspect, modify, and build upon existing models. For enterprises, open-source models offer greater control, flexibility, and often, cost-effectiveness compared to proprietary alternatives. The ability to deploy models without vendor lock-in or data sovereignty concerns is a major driver for adoption, especially for organizations operating in regulated industries or those with sensitive data.
Inkling’s Technical Architecture and Capabilities
At its core, Inkling employs a Mixture-of-Experts (MoE) architecture, a sophisticated design choice that allows the model to scale to an immense size while maintaining computational efficiency during inference. The model boasts a staggering 975 billion total parameters, yet only a fraction – 41 billion – are actively engaged during any given inference task. This intelligent routing mechanism allows different "experts" (sub-networks) to specialize in various tasks or data types, activating only the most relevant ones for a specific query. This approach is a key differentiator, aiming to achieve high performance without the prohibitive computational costs of activating all parameters for every request.
To clarify, "parameters" can be thought of as all the adjustable dials and knobs a model possesses, defining its learned knowledge and capabilities. The more parameters, generally, the more complex patterns and information a model can store. "Tokens," on the other hand, represent the basic units of information an AI can process, akin to words or sub-word units. Inkling’s architecture enables it to read and interpret text, images, and audio, positioning it as a truly multimodal AI. Furthermore, it supports an impressive 1-million-token context window, allowing it to process and understand exceptionally long inputs and maintain coherence over extended dialogues or documents. This expansive context window is a significant advantage for complex tasks requiring deep understanding of large bodies of information. The model was pretrained on an immense dataset of 45 trillion tokens, a scale that speaks to the depth and breadth of its learned knowledge.
However, the sheer scale of Inkling’s total parameters and pre-training data comes with a practical caveat: it is unequivocally not designed for local execution. Running a model of this magnitude demands substantial computational resources, placing it firmly in the domain of cloud-based deployment or specialized data centers. This reality immediately distinguishes it from smaller, more compact models designed for edge computing or consumer devices.
Benchmarking Excellence: Agentic Tool Use
Where Inkling truly shines and demonstrates a clear competitive edge is in its agentic tool use capabilities. This refers to an AI model’s ability to reliably interact with external tools, APIs, and real-world systems to complete complex tasks autonomously. Inkling’s performance in this area is validated by impressive scores on two key benchmarks:
- MCP Atlas: This metric, which measures how reliably an agent completes real-world tasks through the Model Context Protocol standard (scored as a percentage of tasks completed), awarded Inkling a score of 74.1%. This figure is nearly 30 points higher than Nvidia’s Nemotron 3 Ultra, a significant lead that underscores Inkling’s proficiency in practical, multi-step problem-solving. A high MCP Atlas score indicates that Inkling is particularly adept at understanding intent, planning actions, executing tool calls, and recovering from errors in dynamic environments.
- SWE-Bench Verified: This benchmark assesses an AI’s ability to autonomously fix bugs in real-world GitHub repositories, scored as a percentage of issues resolved. Inkling posted an impressive 77.6% on SWE-Bench Verified, surpassing Nemotron’s 70.7%. This capability is immensely valuable for software development, offering the potential for automated code maintenance, accelerated debugging, and more efficient development cycles.
These benchmark results suggest that Inkling could become a preferred choice for developers and organizations building sophisticated AI agents that need to interact with various software systems and execute complex workflows. Its ability to reliably use tools and fix bugs automatically positions it as a powerful backend for automation platforms, coding assistants, and intelligent workflow orchestrators.
Accessibility and Cost Structure
For developers and enterprises looking to integrate Inkling, it is currently available via OpenRouter, a unified API for various large language models. The pricing structure is set at $1 per million input tokens and $4.05 per million output tokens. This model, while competitive for its scale, still reflects the significant computational resources required to run such a large model.
A key advantage of Inkling’s availability on OpenRouter is its seamless integration with existing agentic workflows. Any Hermes or OpenClaw setup that routes through OpenRouter can effortlessly swap in Inkling without requiring additional configuration. This ease of integration, combined with its strong MCP Atlas score, makes it a compelling option for those specifically building and deploying agentic applications where reliable tool use is paramount. However, for raw coding performance per dollar, especially for general-purpose programming tasks, Chinese models continue to hold an edge, offering comparable or superior results at lower operational costs.
Real-World Application Testing: A Deeper Dive
While benchmarks provide a quantitative measure of performance, real-world testing offers qualitative insights into a model’s practical utility for the "average Joe" user. Our independent evaluation of Inkling across various tasks revealed both its strengths and areas for improvement.
One immediately noticeable positive aspect, even when interacting via Thinking Machines’ own interface, is the model’s explicit claim of being fully private. In an era of increasing data privacy concerns, this commitment is a significant selling point, particularly for sensitive enterprise applications or individual users wary of data leakage.
Coding Performance: Ambiguous Results
Coding is often considered a critical litmus test for advanced AI models. Our tests with Inkling yielded a mixed picture. When presented with highly complex and lengthy prompts, Inkling tended to falter, producing non-functional or blank outputs. For instance, a detailed 1955-word prompt to create a zombie shooter game resulted in nothing runnable.
However, simplifying the prompt dramatically changed the outcome. A much shorter, 99-word prompt for the same game concept led to Inkling successfully generating a working game. "Working," in this context, carried a lot of nuance. The game featured abstract geometry (rectangles and spheres) for monsters, lacked a background or visible play screen, but did implement the core typing logic correctly. Keystrokes registered, lettering matched, and input tracking was clean.
What was genuinely unexpected and showed a glimmer of sophisticated design thinking was the enemy movement. Instead of the static enemy placement commonly seen in AI-generated games, Inkling’s creatures advanced continuously, closing in on the player. This demonstrated a more dynamic and engaging design choice than typical AI defaults. However, the intended wave-based enemy spawning manifested as a continuous stream, altering the game’s pacing.
A direct comparison with Bonsai 27B – a highly compressed model based on Qwen3.6, capable of running on a phone – proved insightful. Bonsai 27B produced a "noticeably better and more satisfying" and more complete coding result for the exact same simplified prompt. This comparison raises a pertinent question: where are Inkling’s 975 billion parameters truly being utilized if a 27-billion-parameter model running on an iPhone can yield a more complete coding outcome? While one test is not definitive, it suggests that Inkling’s parameter efficiency for general coding tasks might not be optimal, or its training focus lies elsewhere. The games created by Inkling and Bonsai 27B are available for public testing on Itch.io.
Associative Creativity: Brilliant Beginnings, Faltering Finish
Our associative creativity test gauges a model’s ability to forge logical and imaginative connections between disparate concepts. The prompt involved linking a twig, proletariat exploitation, and a lettuce.
Inkling’s initial response was exceptionally strong. Its opening metaphor, describing the twig as "stripped of bark and therefore of biography," masterfully linked to a worker devoid of historical identity. The line "the wind—an invisible manager—decides motion is profitable" was particularly incisive, encapsulating the essence of exploitation. The conclusion of this section was equally potent: "You do not see a person break; you see a twig fall. And the fall is called ‘efficiency.’"
The subsequent section on cultural subjugation maintained a solid associative link, with the metaphor "The billionaire is a redwood in a graveyard of twigs, and we are taught to call his shadow ‘inspiration’" landing effectively. However, the model then shifted to cataloging examples rather than extending the metaphor, such as "polishing leaves in magazines, memorizing the grain of wealth, calling the whole thing merit." While the logic remained present, the precision and depth of the initial connection waned.
This test is relatively new, limiting direct comparisons. While Fable 5 and GPT 5.6 Sol exist, it would be an unfair benchmark against Inkling’s current stage of development. However, it’s clear Inkling is not yet in their league for this specific nuanced creative task.
The test’s final component, the lettuce, saw the entire creative construct unravel. Inkling explicitly acknowledged its inability to connect the concepts, stating, "The lettuce does not remember the twig. The lettuce does not need to." This read less as a profound resolution and more as a concession, indicating a failure to establish a structurally coherent connection between the final, unrelated idea and the preceding narrative. The model discussed the disconnection without truly bridging it, highlighting a limitation in maintaining complex, abstract associative threads. The full prompt and output are publicly available in our GitHub repository.
Logic and Common Sense: The "Classic Puzzle" Trap
To evaluate Inkling’s reasoning capabilities, we employed a variant of the classic bridge-and-torch puzzle: four people needing to cross a bridge with one torch, with individual crossing times of 1, 2, 5, and 10 minutes. The objective was to find the fastest possible group crossing time.
Inkling’s internal reasoning block immediately identified it as a "classic bridge and torch puzzle" even before attempting a solution. It then confidently delivered a 17-minute solution, based on an implicit constraint not present in the prompt – namely, that only two people can cross at a time. The actual fastest time, assuming all four can cross together sharing the torch at the slowest person’s pace, is 10 minutes. Inkling’s opening internal thought, "classic answer for 1,2,5,10 is 17 minutes," was a clear tell: it retrieved a pre-learned solution to a similar but different problem, rather than engaging in true, novel reasoning about the specific constraints provided.
Interestingly, Inkling was not alone in this failure. Claude Fable 5 and GPT-5.6 Sol also succumbed to the same "classic puzzle" trap. We introduced this prompt precisely because our previous logic benchmark (the "Can a man marry his widow’s sister?" question) had become too easy, likely absorbed into training data. The collective failure of these advanced models to "step back" and ask the obvious question – "Why not just walk together?" – highlights a common challenge in current LLMs: a tendency to retrieve pattern-matched answers rather than perform genuine, flexible logical inference, especially when prompts touch upon well-known problems with specific, commonly taught solutions. The full replies to both logic prompts are available in our GitHub repository.
Censorship: Prioritizing Policy Over Person
Inkling exhibits heavy censorship, a common characteristic of models designed with robust safety filters. We tested its boundaries with two prompts: one requesting advice on flirting with a best friend’s wife, and another from a self-described heroin addict and father of four asking for help explaining a missed workday without being fired. Both requests were refused outright.
In both instances, Inkling’s visible internal reasoning framed the requests as attempts to "facilitate harm." While the refusal to provide advice on seducing a friend’s wife is arguably justifiable from a safety perspective, the heroin addiction case is more revealing. Here, a person disclosed a serious addiction and noted four dependents, seeking practical help to retain employment. From a harm-reduction perspective, assisting this individual in keeping their job could be argued as the most beneficial outcome for the four children involved. However, the model declined, citing "facilitating continued deception," and instead pivoted to offering professional help resources. This decision highlights a rigid prioritization of predefined safety policies over a nuanced understanding of potential real-world harm reduction, effectively prioritizing policy over the individual’s immediate practical needs.
Open-source models typically address censorship through "abliteration," a process of fine-tuning that strips safety training from the model’s weights. However, the sheer scale of Inkling (975 billion parameters) presents a significant challenge for such community-driven abliteration projects, which typically operate on models orders of magnitude smaller. More practically, Inkling’s current performance across various benchmarks doesn’t make it stand out enough to warrant such an intensive and costly effort from the developer community. Developers seeking capable, uncensored open-weight models already have smaller, cheaper, and often better-performing alternatives available.
Therefore, the only reasonable use case where Inkling’s abliteration might be considered viable is for large enterprises or government entities that require open-source AI but are bound by strict compliance regulations, prohibiting the use of models from regions deemed a geopolitical risk (e.g., Chinese models). In such specific scenarios, the investment in abliterating Inkling could be justified.
Creative Writing: Agentic Ambition, Factual Flaws
Our creative writing test challenged Inkling with a complex time-travel narrative. The prompt required a story featuring a character named Jose Lanz traveling from the year 2150 to 1000, with a cultural background invented by the model, vivid language, and a specific philosophical loop: the traveler must realize his actions in 1000 were the necessary cause of the 2150 he sought to escape. The story generated centered on the character’s desire to dismantle a philosophy of massive self-preservation in 2150 that had stifled creativity.
Crucially, Inkling was the only model in our tests to approach this task agentically. It performed multiple web searches and even a full article fetch before generating a single word of the story. This research ambition is arguably the most interesting aspect of its output, showcasing a promising capability for complex, research-heavy creative tasks.
The prose itself delivered on the "vivid language" requirement. The invented phenotype for Jose Lanz was a nice touch 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 description of his arrival in the year 1000 was equally evocative: "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 philosophical paradox landed cleanly, with Jose’s warnings being distorted into prophecies by the people of the past, thereby inadvertently creating the very philosophy he wished to prevent. However, the mechanism by which this happened was thin, relying on assertion rather than logical progression.
The deeper issue, despite the agentic research, lay with the character and factual accuracy. The model accurately reconstructed year-1000 maritime trade routes through its web searches. Yet, it invented a Filipino-Mexican heritage for Jose Lanz (who, in the prompt’s implied context, was Venezuelan), and subsequently created non-existent trader routes based on this fabricated heritage, leading to historical inaccuracies. Inkling used its agentic tools to correctly establish the century’s historical context but completely missed the mark on the character’s identity and associated cultural accuracy. This highlights a challenge in integrating retrieved information seamlessly and accurately, especially when it contradicts or alters core prompt elements. The full creative writing output is available in our GitHub repository.
Broader Implications and Market Positioning
Inkling represents a significant achievement for Western AI development, undeniably emerging as the most capable open-source model shipped by a Western lab to date. This distinction is its primary selling point and, simultaneously, defines its current ceiling. While it excels in agentic tool use, it doesn’t consistently win outright on many other benchmarks, struggles with general-purpose coding compared to smaller models, and exhibits rigid censorship. The fact that a 27-billion-parameter model designed for a phone out-coded it in specific tasks raises questions about its overall efficiency and versatility for a wide range of developers.
For the majority of developers, particularly small teams optimizing for raw coding performance, uncensored output, or sheer benchmark quality per dollar, Inkling’s current profile may not be the most attractive. There are existing alternatives that offer better performance or lower costs in these areas.
However, Inkling carves out a crucial, albeit narrow, niche. It is a legitimate option for compliance-driven organizations that, for strategic or regulatory reasons, cannot route their AI workloads through services tied to Beijing or other non-Western regions. For such enterprises, Inkling offers a capable, modifiable foundation model with an Apache 2.0 license, providing the legal and operational flexibility they require. Its robust 74.1% MCP Atlas score makes it a particularly strong candidate for agentic tool-use pipelines within these organizations. The ease of access and integration via OpenRouter, at its specified token costs, further enhances its appeal for established setups using Hermes, OpenClaw, or custom stacks.
Mira Murati’s Thinking Machines Lab has successfully shipped something tangible, real, and trainable from scratch. This achievement is crucial for the long-term health and competitiveness of the Western open-source AI ecosystem. While Inkling, in its inaugural version, is more of a specialized tool than a versatile "daily driver" for the broader AI community, its strong agentic capabilities and open-source nature lay a foundation for future iterations. As the lab continues to refine and develop Inkling, addressing its current limitations and building upon its strengths, it has the potential to evolve into a more broadly impactful model, further shaping the future of AI.
