Mira Murati’s return to the frontier of AI has finally taken a concrete form. After two years of operating largely out of the spotlight following her departure from OpenAI, she has unveiled Inkling, the inaugural model from her new venture, Thinking Machines Lab. The launch is not just another model drop: Inkling instantly positions itself as arguably the strongest fully open-source, from-scratch large language model developed by a Western team.
The timing matters. Over the past year, Western researchers have been steadily losing ground in the open-source arena. Public leaderboards for open models have become dominated by Chinese systems such as Alibaba’s Qwen, Zhipu AI’s GLM series, and Moonshot AI’s Kimi. Against this backdrop, Nvidia’s Nemotron has been one of the few Western entries to show up in the rankings-yet it has not been seriously regarded as a true state-of-the-art challenger. Inkling lands directly in the middle of this landscape and pushes back on the idea that Western labs can’t compete in fully open releases.
Inkling is available with its full weights released under the Apache 2.0 license and can be downloaded and self-hosted via Hugging Face. That alone is significant: no regional restrictions, no partial checkpoints, and no non-commercial handcuffs. On top of that, it’s exposed through OpenRouter, making it straightforward for developers to test it as a drop-in option alongside other frontier and open models using the same interface.
At the technical level, Inkling adopts a mixture-of-experts (MoE) architecture. The raw parameter count is enormous-975 billion parameters in total-but that doesn’t mean all of them are used on every token. Instead, only about 41 billion parameters are “active” per token, routed through a subset of specialized experts. This design aims to get the benefits of a massive, specialized model without the prohibitive inference cost you’d expect from naively using all 975 billion parameters at once.
One of the most striking early data points is Inkling’s MCP score. MCP (Model Context Performance, or a closely related benchmark concept) is used to evaluate a model’s ability to reason and maintain coherence in extended contexts. On this front, Inkling delivers results that are not just competitive but genuinely impressive, particularly for a model whose weights are fully open and reproducible. It performs at a tier that, until recently, was associated almost exclusively with proprietary frontier systems.
The economic story is more nuanced. While the raw performance-per-token is strong, the price-to-performance calculation is not as straightforward as “cheaper and better than everything else.” The MoE design can reduce effective compute per token, but hosting and serving costs for a nearly trillion-parameter model are still substantial. On commercial endpoints like OpenRouter, Inkling often competes against highly optimized models from other vendors that might be cheaper or more aggressively tuned for latency, even if they fall slightly behind on pure reasoning metrics. For teams evaluating it in production, the calculus isn’t just “Is this the smartest open model?” but “Does the extra performance justify the infrastructure and usage costs for this specific application?”
How Inkling Performs in Practice
In hands-on testing across standard language tasks-reasoning, coding, summarization, and instruction following-Inkling shows several notable strengths.
First, its reasoning capabilities stand out. On complex multi-step prompts, it maintains a clear chain of thought, decomposing problems into parts and then recombining them into coherent answers. This is exactly the sort of behavior that correlates with high MCP-style scores, and it’s critical for workflows like data analysis, research assistance, and technical planning.
Second, Inkling’s coding abilities are competitive with the best open models currently available from the West. It can generate and refactor code in major languages, explain libraries and APIs, and debug nontrivial snippets when provided with error messages or stack traces. While it may not yet outpace the very best proprietary code-focused models in every scenario, for an open system with an Apache 2.0 license, the practical capabilities are compelling.
Third, the model’s instruction following is robust. It generally adheres closely to requested formats, tones, and constraints, which is vital for downstream fine-tuning or building structured tools like agents, chatbots, and workflow automations. The MoE architecture doesn’t appear to have introduced the sort of inconsistency that sometimes shows up in earlier expert-based models; routing appears mature enough that the user doesn’t experience wild swings in quality between similar prompts.
Where It Fits in the Open-Source Landscape
To understand why Inkling matters, it helps to place it within the broader ecosystem of open models. On one side, there are the big proprietary players: closed weights, API access only, and license terms that can change overnight. On the other, there are increasingly capable open-source models-many of them from Asian labs-that have surged up leaderboards and become popular among self-hosters and smaller companies that want maximum control.
Inkling is one of the first Western models trained from scratch that can credibly claim to rival leading open models from Qwen, GLM, and Kimi, while also shipping with a permissive license. For organizations that, for regulatory, legal, or strategic reasons, prefer to rely on Western-developed models, this fills a gap that has been widening for months.
Compared with something like Nemotron, Inkling’s distinguishing feature is not just its scale, but its combination of transparency, licensing, and empirical performance. Nemotron has struggled to be seen as top-tier; Inkling arrives with a clearer technical story and a benchmark profile that looks much closer to modern “best in class” open models.
Cost, Infrastructure, and Deployment Tradeoffs
Even though only a fraction of Inkling’s total parameters are active at any given time, running a 975B-parameter MoE model is still not trivial. For developers and companies considering self-hosting, this raises some practical questions:
– Hardware requirements: You will likely need multiple high-memory GPUs or a cluster with a sharded serving setup to achieve acceptable latency and throughput. This is not a “runs comfortably on a single consumer GPU” model.
– Latency vs. quality: MoE routing adds some complexity to inference. While the effective active parameter count is roughly in line with a ~40B dense model, routing, loading, and expert management can affect responsiveness. For interactive applications, careful tuning and caching will be needed.
– Cloud vs. local: For many, using Inkling via a routing platform like OpenRouter or a managed inference provider will make more sense than building and maintaining infrastructure in-house. The tradeoff is recurring usage cost versus upfront hardware and maintenance.
All of this feeds into the price-to-performance puzzle. If your workload is highly sensitive to response quality-complex planning, research, or specialized expert assistance-Inkling’s extra reasoning depth may justify the bill. If your primary need is low-cost, high-volume, semi-structured text generation, a smaller, cheaper dense model (or a compressed variant of Inkling, if and when those appear) might be a better choice.
Who Should Consider Using Inkling?
Inkling is not designed as a casual toy; it targets serious users who care about model sovereignty, quality, and flexibility:
– AI startups and product teams who want to build on top of a strong base model without being locked into a single vendor’s API or pricing. With Apache 2.0 licensing, you can fine-tune, embed, and even resell systems built on top of Inkling without complicated legal entanglements.
– Enterprises that have data-governance or geopolitical constraints and prefer open, auditable models originating from Western labs. Inkling offers them a middle ground between closed proprietary systems and foreign open models.
– Researchers and tool builders who want to experiment with large MoE architectures, interpretability, or alignment methods on a modern-scale system. Full weights and permissive terms are invaluable for this kind of work.
– Developers seeking a “best available” Western open-source base model for custom assistants, coding copilots, or domain-specific agents.
If your priority is maximum control, legal clarity, and customization potential-and you’re prepared for the hardware and cost implications-Inkling is one of the most attractive options now on the table.
Strengths and Limitations
Like any model, Inkling has areas where it shines and others where it shows the typical weaknesses of current LLMs.
Strengths:
– High-quality reasoning and long-context performance.
– Strong instruction following and formatting fidelity.
– Competitive coding assistance for an open model.
– Fully open weights, with a permissive license suitable for commercial use.
– Trained from scratch by a Western lab, providing an alternative to the growing dominance of non-Western open models.
Limitations:
– Heavy infrastructure footprint, even with MoE efficiency.
– The economics may be less favorable for workloads that don’t strictly need frontier-tier reasoning.
– As with all large LLMs, it can hallucinate, misinterpret ambiguous prompts, or produce confidently wrong answers.
– Safety and alignment are reasonably good but not infallible; high-risk use cases still require additional layers of filtering and control.
How Inkling Compares to Other Open Models
For teams trying to decide whether Inkling should be their default open model, it’s useful to contrast it with some common alternatives:
– Against smaller dense models (7B-34B): Inkling provides significantly better reasoning, instruction adherence, and robustness under complex prompts. However, smaller models are easier and cheaper to run, making them attractive for lightweight applications and edge deployments.
– Against leading Chinese open models: Inkling closes much of the gap that had opened up over the past year. While head-to-head benchmarking will continue to evolve, it offers similar or better performance in several categories, with the added benefit of Western provenance and a well-understood legal environment.
– Against proprietary frontier models: The very top-tier closed models still have an edge in some combination of reasoning, multimodality, and tool use. But for many practical use cases, the difference is small enough that the openness and flexibility of Inkling become more important than squeezing out the last few percentage points of benchmark performance.
Future Directions and Ecosystem Potential
The release of Inkling is likely just the starting point for an ecosystem. Expect:
– Instruction-tuned and domain-specialized variants fine-tuned for coding, legal, medical, or enterprise tasks.
– Smaller distilled models that capture much of Inkling’s behavior in a more compact form suitable for cheaper inference.
– Tooling and libraries optimized around MoE inference, routing, and serving, making it easier to deploy at scale without bespoke engineering.
– Research derivatives focusing on interpretability, safety, and efficiency, taking advantage of the fully open weights and training recipe.
If Thinking Machines Lab continues to iterate at this level, Inkling could become the foundation for a family of models that serve different segments of the market, from high-performance cloud deployments down to resource-constrained environments via distillation.
Is Inkling Really the “Best Open-Source Model in the West”?
The claim that Inkling is the best open-source model trained from scratch by a Western lab is defensible under current conditions. Competitors either lag on core performance metrics, are not trained fully from scratch, or impose more restrictive terms. With its MoE architecture, impressive MCP-style performance, and Apache 2.0 license, Inkling sets a new bar in this specific slice of the ecosystem.
However, “best” depends heavily on your priorities. If you care above all about:
– Performance under complex reasoning and long context: Inkling is near the top of the open Western stack.
– Ease and cost of deployment: You might favor smaller dense models or cloud-only proprietary models.
– Total ecosystem maturity: Some existing models have more tooling and fine-tuned variants today, though that gap will likely close quickly.
Bottom Line
Inkling represents a meaningful shift in the balance of power in open-source AI. After a period in which Western labs appeared content to let others dominate the open leaderboard space, Mira Murati’s new model signals a renewed willingness to compete at the highest level-openly, with full weights and permissive licensing.
For developers, researchers, and companies seeking a powerful, Western-built, fully open LLM, Inkling is currently one of the strongest options available. Its mixture-of-experts design, impressive benchmark performance, and flexible licensing together justify the growing attention it’s receiving, even if its true economic sweet spot will depend on how each user weighs cost, infrastructure, and quality.