Thinking Machines, the AI lab founded by former OpenAI CTO Mira Murati, has released a second model called Inkling Small, shifting attention from raw scale to efficiency. According to coverage from The Decoder, the new model uses far fewer active parameters than the company’s earlier Inkling model while coming close to it on a broad benchmark index and outperforming it on some coding and reasoning tests.

That positioning matters because it speaks directly to a pressure point in the current model market: many builders and enterprise buyers want strong reasoning performance without paying the inference costs, latency, and deployment complexity that often come with larger frontier systems. If Thinking Machines can show that a smaller open model can stay competitive on meaningful tasks, it strengthens the case for fine-tuned, domain-specific deployments rather than a one-size-fits-all race toward ever-larger models.

A smaller model with near-parity ambitions

The core news is straightforward. Thinking Machines has introduced Inkling Small as an open-weights reasoning model, and The Decoder reports that it is less than a third the size of Inkling on the key parameter figures being cited. Based on benchmark tracking from Artificial Analysis, Inkling Small scores 40 on the firm’s Intelligence Index, compared with 41 for Inkling.

The same reporting says the model has 276 billion total parameters and 12 billion active parameters. The distinction matters. In current model design, especially for sparse or mixture-style systems, the active-parameter count can be a better proxy for runtime cost than total parameters alone, because it reflects how much of the network is actually engaged on a given token. While the available source material does not detail the architecture behind Inkling Small, the gap between total and active parameters suggests Thinking Machines is emphasizing efficiency not just in model marketing but in the way inference is likely being managed.

VentureBeat, in a separate report, framed the launch similarly, describing Inkling Small as an open source AI model that nears the performance of its predecessor at roughly a quarter of the size. Because the full text of that report was not available in the source evidence, the more detailed technical and benchmark descriptions here rely primarily on The Decoder’s reporting and the third-party benchmark references it cites.

Where Inkling Small appears stronger

According to The Decoder, Inkling Small beats its larger sibling on several reasoning and coding-oriented evaluations. The reported examples include Humanity's Last Exam, where Inkling Small scored 32% versus 30% for Inkling, and GPQA Diamond, where it posted 89% versus 87%.

Those are not trivial wins. Humanity's Last Exam has been used as a difficult benchmark designed to probe more advanced reasoning ability, while GPQA Diamond is commonly cited in discussions of high-end scientific and expert-level question answering. Gains there suggest Thinking Machines may have tuned Inkling Small to perform well on compact, difficult reasoning tasks rather than merely preserving broad baseline capability.

At the same time, the model reportedly trails the larger Inkling on agent-based tasks and factual knowledge. That tradeoff is important. For teams building coding assistant products, internal research copilots, or narrow reasoning workflows, better performance per token can be more valuable than broad knowledge depth. But for general-purpose agents that must retrieve facts, use tools over longer horizons, or operate reliably across many enterprise tasks, those deficits could still matter.

The other performance angle is output efficiency. The Decoder reports that Inkling Small averages 24K output tokens per task, compared with 45K for Deepseek V4 Flash and 78K for GPT-5.4 mini. If that comparison holds under similar conditions, it points to a practical advantage: a model that reaches a result with fewer generated tokens can lower cost and reduce user-visible latency. As with many benchmark-style efficiency comparisons, however, the exact task mix and evaluation setup matter, and buyers should avoid reading too much into a single token-efficiency number without independent testing.

Open-weights release and product packaging

The release is also notable for how Thinking Machines is packaging the model. The Decoder says Inkling Small supports text, image, and speech inputs, giving it a multimodal profile rather than positioning it as a text-only reasoning engine. It also has a 256K-token context window, which places it in the range expected for large-document analysis, codebase work, and extended conversational sessions.

On access, the model is being distributed under Apache 2.0, with weights available on Hugging Face. That licensing choice is significant for commercial adoption. Apache 2.0 is generally viewed as enterprise-friendly because it supports broad use, modification, and distribution without some of the legal uncertainty that can accompany more restrictive community licenses.

Thinking Machines is also making Inkling Small available through Tinker Playground, where users can fine-tune the model in the browser, according to The Decoder. That fits a product strategy centered less on selling one canonical model endpoint and more on turning the base model into a customizable starting point. The company, as described in the report, is positioning its models as a foundation that users can adapt with their own data.

That message lands in a market increasingly interested in ownership and specialization. Many product teams now want control over latency, deployment topology, safety behavior, and domain adaptation. An open-weights model distributed through Hugging Face and paired with lightweight tuning tools gives Thinking Machines a clearer route into that buyer segment than a closed API-only offering would.

What the evidence supports — and what it does not

The evidence behind Inkling Small’s launch is promising, but still relatively thin. The most detailed source in this cluster is The Decoder, which in turn attributes the main performance framing to Artificial Analysis. That means the headline claim — near-Inkling capability at much smaller active size — is grounded in third-party benchmark tracking rather than in a full technical paper included in the provided evidence.

Some claims are straightforward and likely easy to verify directly, including the Apache 2.0 license, the presence of weights on Hugging Face, the 256K context window, and availability via Tinker Playground. Others require more caution. Benchmark scores such as the Intelligence Index, Humanity's Last Exam, and GPQA Diamond depend on evaluation protocols, prompting choices, and model versions. The claim that no open model of equal or smaller size scores higher comes from Artificial Analysis as quoted by The Decoder, not from an independently reproduced market-wide audit in the material provided here.

There is also no detailed deployment data in the source evidence. We do not yet have direct numbers on throughput, real-world serving costs, enterprise usage, fine-tuning stability, or safety performance under adversarial testing. So while Inkling Small looks compelling on paper as an efficient open-weights release, buyers should treat the current picture as benchmark-led rather than field-proven.

Why this launch matters for builders and enterprises

For AI builders, the strongest signal in this release is not simply that Thinking Machines shipped another model. It is that the company is arguing for a different optimization target. In a market still dominated by flagship model announcements, Inkling Small suggests there is room for a product strategy built around useful reasoning at lower active size, lower token consumption, and easier customization.

That could resonate in several workflows. A coding assistant that needs fast iterative suggestions may benefit more from token efficiency than from the broadest possible world knowledge. Internal enterprise AI systems built around proprietary documents may prefer a model that can be fine-tuned on company data and deployed under a permissive license. Teams handling multimodal inputs may see value in one model that can process text, images, and speech without defaulting to a larger proprietary stack.

The competitive angle is also clear. By comparing output-token efficiency against Deepseek V4 Flash and GPT-5.4 mini, the launch enters a crowded middle tier of models competing on cost-performance, not just absolute benchmark leadership. That is where many enterprise AI budgets are actually decided. The question is less “what is the smartest model available?” and more “what is good enough, controllable enough, and affordable enough to put into production?”

For Thinking Machines, this release helps define its identity. Rather than trying to outmuscle the biggest closed-model vendors on spectacle alone, the company appears to be betting that open deployment, efficient reasoning, and fine-tuning flexibility can carve out a durable position.

What to watch next

The next signals to monitor are practical rather than promotional. First, independent testing of Inkling Small on coding assistant tasks, agent reliability, and factual grounding will show whether the benchmark story translates into production use.

Second, watch whether the Hugging Face release generates meaningful community traction: derivative checkpoints, tuning recipes, framework integrations, and reproducible latency reports. Open-weights launches matter more when they attract an ecosystem, not just downloads.

Third, the role of Tinker Playground will be worth tracking. If browser-based fine-tuning lowers the barrier to specialization, Thinking Machines could gain adoption among smaller teams that lack heavy ML infrastructure.

Finally, any future technical disclosures from Thinking Machines will matter. More detail on architecture, training methods, safety behavior, and inference economics would help enterprises assess whether Inkling Small is mainly a benchmark-efficient model or a genuinely operational alternative to larger proprietary systems.

Creati.ai perspective

Inkling Small fits a broader shift in enterprise AI buying: optimization is moving from raw model prestige toward deployment math. A model that is slightly worse on a leaderboard but materially better on token use, tuning flexibility, and licensing can be the more important product.

The open question is whether Thinking Machines can turn that thesis into adoption. Inkling Small gives the company a credible entry in the open-weights market, especially through Apache 2.0, Hugging Face, and Tinker Playground. But to break through, it will need more than benchmark proximity to Inkling. It will need evidence that teams can reliably build real enterprise AI and AI agents on top of it with better cost, control, and speed than closed alternatives.