The Verbosity Paradox: How Thinking Machines’ Inkling 2 Small is Redefining Agentic Value
The landscape of open-weight AI has long been defined by a struggle between raw intelligence and deployment efficiency. For Western developers, the choice has often been a binary one: opt for the massive, closed-source proprietary giants or settle for smaller, highly efficient models that lack the "
The landscape of open-weight AI has long been defined by a struggle between raw intelligence and deployment efficiency. For Western developers, the choice has often been a binary one: opt for the massive, closed-source proprietary giants or settle for smaller, highly efficient models that lack the "reasoning" depth required for complex, autonomous tasks.
That binary is being challenged. Thinking Machines, the startup led by former OpenAI CTO Mira Murati, has officially entered the fray with the release of Inkling 2 Small. While the model is currently accessible via their proprietary Tinker platform, the company has signaled a massive commitment to the open-source community, promising to release the model weights under an Apache 2.0 license once safety testing is complete.
Inkling 2 Small is not just another incremental update; it is a high-performance base model designed specifically to be the foundation for specialized, fine-tuned applications. However, its arrival brings with it a fascinating technical paradox: it is exponentially more capable at agentic tasks, yet it achieves this intelligence through a massive increase in token verbosity that fundamentally alters the economics of AI interaction.
The Architecture of Intelligence: MoE at Scale
At its core, Inkling 2 Small utilizes a Mixture-of-Experts (MoE) architecture. This design allows the model to maintain a massive total parameter count of 594 billion, while remaining computationally efficient during inference by only activating 17 billion parameters per specific task. This "sparse" approach is what allows the model to punch significantly above its weight class in terms of reasoning capabilities without requiring the energy footprint of a dense 600B+ parameter model.
The training regimen behind Inkling 2 Small suggests a surgical approach to data. Rather than simply throwing more compute at a larger dataset, Thinking Machines utilized a relatively compact training set of 12 trillion tokens. The real "secret sauce," however, lies in the reinforcement learning (RL) phase. The team dedicated 3 trillion tokens to an intensive RL process, utilizing a cluster of 2,048 GPUs over a nine-day period. This phase was specifically engineered to optimize for long-horizon coding tasks—the kind of multi-step logical reasoning that separates a simple chatbot from a functional AI agent.
Benchmarking a New Contender
The performance metrics for Inkling 2 Small suggest it is already a heavyweight in the Western open-weight space. On the Artificial Analysis Intelligence Index, the model earned a score of 31, placing it on par with Google’s Gemini 3.8 Flash.
While the raw intelligence score is impressive, the most dramatic leap is seen in agentic performance. On the AutomationBench-AA, Inkling 2 Small saw its score skyrocket from a mere 5% in previous iterations to a staggering 46%. This indicates that the model is no longer just predicting the next word; it is beginning to understand how to execute multi-step workflows.
Multimodal and Visual Reasoning
Inkling 2 Small is natively multimodal, capable of processing both text and image inputs. This isn't just a secondary feature; it is a core competency. The model shows leadership among open-weight competitors in several vision-centric benchmarks:
- Document Reasoning (GDP.pdf): Superior ability to parse complex layouts and extract structured data.
- Multimodal Academic Tasks (MMMU Pro): High proficiency in solving complex problems that require visual-spatial reasoning.
- Infographic Interpretation: A notable ability to translate visual data into coherent textual analysis.
The Coding Frontier
In the realm of software engineering, the model is a powerhouse, scoring 86.3% on SWEBench Pro v2. While it still trails behind Deepseek-V4.1-Flash in certain niche coding benchmarks, its ability to handle complex, long-context programming tasks makes it a formidable candidate for autonomous coding agents.
The Efficiency vs. Cost Paradox
Herein lies the central tension of the Inkling 2 Small release. While the model is "efficient" in its parameter activation, it is "inefficient" in its communication style.
To achieve its high reasoning scores, the model has become significantly more verbose. Inkling 2 Small produces roughly four times more output tokens than its predecessor. This increased "thinking" time—manifested as more text—creates a complex economic profile.
On one hand, the cost per task is remarkably low. At $1.18 per task, it is a fraction of the cost of Google’s Gemini 3.8 Flash ($7.77) or Claude Sonnet 5.5 ($10.02). This pricing structure puts it in direct competition with low-cost models coming out of China. On the other hand, the six-fold increase in cost per task compared to its previous version is a direct result of that massive token output.
Users are essentially paying for the "thought process." You are getting a much smarter agent, but you are paying for every single "thought" the model writes down.
Known Limitations and Trade-offs
No model release is without its scars, and Inkling 2 Small is no exception. The pursuit of higher intelligence has come at the cost of several operational nuances:
- The Response Rate Drop: While hallucination rates have decreased, the overall response rate has fallen from 77% to 53%. This suggests that the model’s increased reasoning threshold causes it to "refuse" or fail to answer questions it deems outside its high-confidence reasoning bounds.
- Communication Inefficiency: The model is prone to "over-explaining," which can be frustrating for users seeking concise answers.
- Lack of Audio Support: Unlike some of its contemporaries, Inkling 2 Small remains strictly text and image-based.
- Optimization Gaps: The model is currently not fully optimized for casual chat, appearing more like a specialized engine than a conversationalist.
- Safety Overreactions: There are noted instances where the model overreacts to sensitive health-related topics, potentially limiting its utility in medical or wellness-adjacent applications.
Conclusion: A Strategic Pivot for Open Weights
Inkling 2 Small represents a strategic shift in how Western AI companies approach the "open" market. Rather than trying to match the sheer scale of closed-source models, Thinking Machines is focusing on the utility of the model—specifically its ability to act as an agent.
By prioritizing agentic reasoning and multimodal understanding, and by offering a roadmap toward an Apache 2.0 release, Thinking Machines is positioning itself as the primary alternative for developers who need high-tier intelligence without being locked into a proprietary ecosystem. The "verbosity tax" is a real concern, but for those building the next generation of autonomous software, the trade-off of more tokens for more intelligence may be a bargain worth making.
