Mira Murati’s Inkling: A Deep Dive Into the West’s Most Ambitious Open‑Source AI Model
After stepping away from OpenAI, Mira Murati disappeared from the spotlight for almost two years. That silence has now been broken with Inkling, the inaugural large language model from her new venture, Thinking Machines Lab.
Inkling isn’t just another entry in the model zoo. It is, at this moment, arguably the strongest fully open‑source model that has been trained from scratch by a Western research lab-both technically and symbolically important in a landscape that has been tilting toward Asian players in open models.
The model is already accessible via OpenRouter, with full weights published under the permissive Apache 2.0 license. That combination-serious performance, a powerful architecture, and genuinely open licensing-instantly puts Inkling on the radar for developers, startups, and enterprises that are wary of closed, policy‑shifting APIs.
Below is a structured look at what Inkling actually is, how it performs, where it shines, and where it still raises questions.
—
A Western Counterpunch in the Open‑Source Race
Over the past year, the open‑source leaderboard has been dominated by Chinese labs and regional players:
– Alibaba’s Qwen family
– Zhipu AI’s GLM series
– Moonshot AI’s Kimi models
By contrast, Western labs have struggled to place competitive, fully open models near the top. Nvidia’s Nemotron has been one of the few Western entries visible on strong benchmark tables, but even supporters rarely call it truly “state of the art” when stacked against the newest releases from Asia.
Inkling changes that dynamic. It arrives as:
– A Western‑built, from‑scratch model (not a finetune of someone else’s weights)
– Released with full weights
– Licensed under Apache 2.0-one of the least restrictive licenses available
This combination instantly makes Inkling a flagship for Western open‑source AI efforts, setting a new bar others will be measured against.
—
Architecture: Mixture‑of‑Experts at Serious Scale
At the heart of Inkling is a mixture‑of‑experts (MoE) architecture. Instead of running a single monolithic set of parameters on every token, MoE models selectively activate a subset of “experts” for each input.
Key architecture facts:
– Total parameters: 975 billion
– Active parameters per token: 41 billion
In other words, Inkling has nearly a trillion parameters in its overall “brain,” but for any given operation, only about 41 billion are actually in play. This gives it:
– The expressive capacity of an ultra‑large model, because you have many specialized experts
– Closer to 40B‑class runtime behavior, because you don’t activate the entire parameter pool at once
This design is particularly attractive for:
– Large‑scale inference clusters that want high throughput with powerful reasoning
– Organizations that can tolerate moderate hardware demands but want top‑tier performance
– Future fine‑tuning regimes, where some experts can be adapted for domain‑specific tasks without disturbing the entire network
MoE models are complex to train and tune, but when done well, they often deliver a better performance‑per‑FLOP profile than dense models of similar raw size. Inkling appears to take full advantage of this.
—
Licensing and Availability: Apache 2.0, Real Openness
Inkling’s release under Apache 2.0 is more than a legal footnote-it is a strategic stance. Apache 2.0 allows:
– Commercial use without revenue‑sharing obligations
– Proprietary integrations and products
– Modification and redistribution of the model and derivatives
Technically, the model is:
– Distributed with full weights
– Unencumbered by regional or “for research only” restrictions
– Already integrated into routing layers like OpenRouter, making it easy to test and compare against competitors
For companies that have been hesitating to commit to models under more constrained or ambiguous licenses, Inkling represents a safe, high‑ceiling choice.
—
MCP Performance: A Genuinely Standout Score
On standardized multi‑capability benchmarks (MCP‑style composite scores that combine reasoning, coding, language understanding, and safety), Inkling posts results that are not just respectable, but genuinely impressive in the context of open models.
While exact numbers depend on the benchmark suite and configuration, early testing suggests:
– Strong general reasoning for an open model
– Competitive performance against leading Western options
– A surprisingly narrow gap to some proprietary frontier models on structured tasks
This matters for teams that rely on objective leaderboards rather than marketing claims. Inkling isn’t merely “good for open‑source”-it’s solid in absolute terms on multi‑task capability scores.
—
Price‑to‑Performance: Where Things Get Complicated
Raw capability is only half the story. The economics of using Inkling are more nuanced:
1. MoE benefits
– You get high effective capacity (975B total) with lower active computation (41B), improving cost‑effectiveness versus an equally strong dense model.
2. But not a budget toy
– A 41B‑active‑parameter model is still heavy.
– Compared to compact 7B-14B models, inference remains more expensive per token, even if it delivers better output quality.
3. Cloud vs self‑hosted trade‑offs
– Via hosted providers, you’re paying for GPU time, bandwidth, and orchestration.
– On your own hardware, you face capital expenditure, engineering complexity, and ongoing operations.
4. Use‑case sensitivity
– For high‑value tasks (e.g., code generation, complex reasoning, legal drafting), a more capable model like Inkling may be economically superior despite higher per‑token costs-fewer retries, less editing, better automation.
– For simple chatbots, FAQ bots, or large‑scale low‑stakes content generation, smaller models will often win on total cost of ownership.
In short, Inkling’s price‑performance looks excellent if you genuinely use its capabilities. If your workload is trivial, you’re overbuying.
—
Real‑World Testing: How Inkling Behaves
In practical evaluation across common workloads, Inkling shows a distinct profile.
1. General conversation
Inkling handles open‑ended dialogue with:
– Coherent, context‑aware responses
– Good memory over medium‑length exchanges
– Reasonably natural tone, without the “flat,” robotic style of some smaller open models
It doesn’t feel like an early open‑source project. It feels like a modern assistant capable of handling multi‑step instructions.
2. Reasoning and problem‑solving
On structured reasoning tasks (multi‑step math, logic puzzles, analytical questions), Inkling:
– Maintains chains of thought over longer sequences
– Handles instructions and sub‑goals without losing context as quickly as many 7B-30B models
– Provides explanations that often show genuine intermediate reasoning, not just pattern‑matching answers
This is critical for enterprise deployments where explainability and reliability matter more than just text fluency.
3. Coding and technical tasks
Inkling appears particularly competitive as a coding model:
– Understands multi‑file project structures when provided with enough context
– Generates reasonably idiomatic code in popular languages
– Handles refactoring and debugging requests with usable suggestions
While it may not yet dethrone the best closed‑source code models, it is strong enough to power in‑house developer tools without licensing headaches.
4. Multilingual capabilities
Inkling is primarily tuned for English, but like most large models of this class, it can operate in several major languages with varying quality. For global deployments, localized fine‑tuning or tailored adapters would likely be needed to reach top‑tier fluency outside English.
—
How It Stacks Up Against Other Open Models
Putting Inkling in context:
– Versus Qwen / GLM / Kimi
– Those models still set the pace on many open leaderboards.
– However, they are developed and governed under very different regulatory and geopolitical conditions.
– Inkling offers a Western‑governed alternative with comparable ambition and a highly permissive license.
– Versus Western open families (e.g., Nemotron)
– Nemotron helped demonstrate that Western labs could at least show up in serious benchmarks, but it was not widely perceived as the new gold standard.
– Inkling, by contrast, is a deliberate attempt to reclaim the top end of the open‑source spectrum for Western research.
– Versus smaller community models
– 7B-14B models remain more accessible for hobbyists and smaller teams.
– Inkling is more of a “flagship” model-stronger capabilities, but with hardware and cost expectations to match.
For organizations that want a high‑end open model from a Western lab, Inkling is now the obvious reference point.
—
Deployment Considerations: Who Should Actually Use Inkling?
Inkling isn’t a one‑size‑fits‑all solution. It’s best suited for:
– Enterprises building serious products
Where quality, reliability, and future‑proof licensing outweigh minimal inference cost.
– Startups in regulated or sensitive domains
That need control over weights, the ability to self‑host, and clear legal terms for data governance and compliance.
– R&D teams and AI infrastructure providers
Looking to experiment with MoE architectures at scale, optimize serving pipelines, or build custom expert layers on top.
On the other hand, it may be excessive for:
– Simple marketing copy farms
– Lightweight chatbots with millions of low‑value interactions
– Edge deployments with tight compute budgets
Those use cases are better served by small, efficient models, possibly fine‑tuned for the specific domain.
—
Strategic Impact: Why Inkling Matters Beyond Benchmarks
Inkling’s launch is as much a signal as it is a product release. It shows that:
1. Western labs are still willing to train very large, fully open models from scratch, not just derivative finetunes.
2. Mixture‑of‑experts is now a mainstream choice for high‑end open models, not just a research curiosity.
3. Permissive licensing can coexist with strong performance, pushing back on the idea that serious models must be locked behind proprietary APIs.
This has downstream consequences:
– Other Western players are now under pressure to either match Inkling’s openness or explain why they won’t.
– Enterprises that have been holding back from committing to open models now have a realistic top‑tier option to evaluate.
– The geopolitical balance of open‑source AI becomes less lopsided, which may matter for policy, regulation, and long‑term AI governance.
—
Limitations and Open Questions
Despite its strengths, Inkling isn’t immune to caveats:
– Hardware demands remain substantial
Even with MoE efficiency, serving a 41B‑active‑parameter model at scale requires serious GPU resources and engineering skill.
– Safety and alignment are still imperfect
Like any large model, Inkling can hallucinate, reflect bias in training data, or produce unsafe content if deployed without proper guardrails or moderation layers.
– Ecosystem maturity is just beginning
Tooling, adapters, fine‑tune checkpoints, and best‑practice playbooks will take time to develop around Inkling specifically. Early adopters will be blazing the trail.
– Comparative leadership is dynamic
Being the strongest Western open model today doesn’t guarantee that status in six or twelve months. The open‑source race moves fast, and new entrants will inevitably appear.
—
How to Decide if Inkling Fits Your Stack
When evaluating whether to adopt Inkling, teams should walk through a few practical questions:
1. Do you truly need frontier‑level reasoning and coding ability, or will a smaller model suffice?
2. Are you prepared to manage the infrastructure needed to run or fine‑tune a model of this scale, or will you rely on hosted providers?
3. Is a permissive license a key requirement for your long‑term business plans (white‑label products, on‑prem deployments, highly regulated customers)?
4. How critical is model governance, auditability, and the ability to inspect or modify weights?
If the answers point toward control, capability, and long‑term independence from closed APIs, Inkling becomes a compelling candidate.
—
Verdict: A New Benchmark for Western Open‑Source AI
Inkling is not a marketing stunt or a minor incremental release. It is a serious, technically ambitious mixture‑of‑experts model-975B parameters total, 41B active-that delivers genuinely strong multi‑capability performance, with full weights available under Apache 2.0.
Western labs have been trailing in the open‑source race, while Qwen, GLM, and Kimi dominated the leaderboards and Nvidia’s Nemotron struggled to be seen as truly top‑tier. Inkling changes that narrative. It gives Western developers and enterprises a high‑end, from‑scratch open model that can anchor real products and serious research.
Its price‑to‑performance profile is subtle: not cheap in absolute terms, but compelling when applied to high‑value workloads that leverage its strength in reasoning, coding, and complex instruction following. For trivial tasks, it will be overkill; for ambitious AI systems, it may be exactly the level of model many teams have been waiting for.
In the broader story of open‑source AI in the West, Inkling is a milestone-and very likely the new reference point future releases will be measured against.