Inkling 2 Small: A 594B MoE Model That Can "See" and Revise

Title

Inkling 2 Small: A 594B MoE Model That Can "See" and Revise

Meta Description

Thinking Machines Lab releases Inkling 2 Small, a 594B-parameter MoE model that activates only 17B, focused on long-horizon coding and visual tasks where it can see and refine its own output, with inference and fine-tuning available via the Tinker platform.

Article Content

Thinking Machines Lab released Inkling 2 Small on Product Hunt, a model built around the idea of "being able to see and refine what it builds." It uses a 594B-parameter MoE architecture that activates only 17B during inference, targeting long-horizon coding, visual tasks, and agentic scenarios, with inference and fine-tuning services offered through the Tinker platform. It topped Product Hunt on launch day with 100 votes.

Model Positioning: From "Generation" to "Self-Correction"

The tagline of Inkling 2 Small, "A model that can see and refine what it builds," captures its core difference: it doesn't just output code or visual content, it can observe its own output and iterate on it. This points to a key direction in the current competition over model capabilities—closed-loop refinement. Traditional models tend to generate results in one shot, requiring users to manually verify and adjust; Inkling 2 Small attempts to complete the "generate-observe-improve" loop internally, making it especially suited to long-horizon coding tasks and scenarios that require visual feedback.

In terms of parameter design, the MoE architecture with 594B total parameters and 17B activated strikes a balance between capacity and inference cost. The large total parameter count ensures knowledge storage and expressive capability, while the low activation count controls the compute cost of a single inference pass. This design is not uncommon in recent large-model releases, but Thinking Machines Lab ties it to "visual feedback" and "long-horizon tasks," forming a differentiated positioning.

Closed-loop refinement refers to a model's ability, after generating an initial output, to observe the result through some feedback mechanism (such as code execution errors, rendered screenshots, interface states, etc.) and iterate and improve accordingly, forming a "generate-observe-revise" cycle. This is fundamentally different from traditional one-shot generation: the latter relies on users to manually verify and adjust, whereas closed-loop refinement internalizes part of the verification and optimization process into the model's workflow. Implementing closed-loop refinement typically requires the model to have multimodal perception capabilities (such as reading screenshots and understanding visual layouts) as well as long-context memory (remembering prior modification history), and may combine external tools (such as code executors and browser rendering engines) to obtain feedback signals. In long-horizon coding tasks, this capability helps reduce manual intervention, but it is technically complex to implement, and the quality and latency of feedback signals affect actual results.

Architecture and Capabilities: MoE Supports Long-Horizon Tasks

The MoE (mixture of experts) architecture lets the model invoke only part of its parameters during inference, thereby reducing latency and cost while maintaining large-model capabilities. Inkling 2 Small's 17B activated parameters mean that the scale actually involved in computation per forward pass is far smaller than 594B, which is especially important for agentic tasks that require multiple rounds of iteration—agents often need to repeatedly call the model, observe the environment, and adjust strategies, and inference costs accumulate with each round.

The official description's mention of "long-horizon coding" and "visual tasks" suggests the model has been specifically optimized for code generation and visual understanding. Long-horizon coding tasks typically involve multi-file, multi-step modifications, requiring the model to remember context and maintain consistency; "being able to see and refine" may mean the model can read screenshots, rendered results, or interface states and revise code or design accordingly. This capability has direct value for scenarios such as front-end development, UI automation, and game development.

The core idea of the MoE (mixture of experts) architecture is to divide the model's parameters into multiple "expert" subnetworks and introduce a learnable routing network. During each forward pass, the routing network dynamically selects and activates a small number of experts (usually top-k, such as top-2) based on the input content, rather than invoking all parameters. This way, the model's total parameter count can be made very large to enhance knowledge capacity, while the number of parameters actually involved in computation (activated parameters) can remain relatively low, significantly reducing the compute cost and latency of a single inference pass. For example, of Inkling 2 Small's 594B total parameters, only 17B are activated each time, meaning only about 2.9% of the parameters participate in computation. This design is particularly well suited to agentic scenarios that require repeated model calls, because the cumulative cost of multiple rounds of iteration becomes more manageable. However, MoE also brings challenges such as training stability, load balancing, and memory footprint, and the quality of the routing strategy directly affects model performance.

Platform and Ecosystem: Tinker Provides Inference and Fine-Tuning

Inkling 2 Small offers inference and fine-tuning services through the Tinker platform. Tinker is Thinking Machines Lab's model service platform, where users can call the model directly or fine-tune it on their own data. For enterprise users and developers, fine-tuning capability is key to deployment—general-purpose models rarely cover all vertical scenarios, and fine-tuning allows the model to adapt to specific codebases, design specifications, or business processes.

The Product Hunt page shows Mira Murati and Tianyi Zhang as the makers. Mira Murati previously served as CTO at OpenAI, and developments at her new company Thinking Machines Lab have attracted ongoing attention. Inkling 2 Small is the company's productized debut to the developer community, and choosing Product Hunt for the launch also indicates its desire to reach early developers and technical decision-makers.

Market Response and Competitive Position

A launch-day result of 100 votes and first place is a decent start on Product Hunt, but with only 2 comments, the discussion has yet to fully develop. Product Hunt votes reflect more the interest of early adopters; the real test lies in retention and word of mouth after developers actually use it.

From a competitive standpoint, Inkling 2 Small enters a crowded field: long-horizon coding and agentic tasks are already key focus areas for multiple model vendors. Its differentiation lies in the "visual feedback loop" and the cost advantages of the MoE architecture. But whether it can perform reliably in real tasks still requires more third-party evaluations and user case studies for verification. For developers, what's worth watching is the actual inference cost on the Tinker platform, the barrier to fine-tuning, and the model's performance on complex codebases and visual tasks.

A Few Questions Worth Watching

First, where the configuration of 594B total parameters and 17B activated stands among MoE models—the company has not provided a direct comparison with competitors. Second, the specific implementation of "being able to see and refine" is not elaborated on the Product Hunt page; whether it is a built-in visual encoder or multimodal input requires further review of technical documentation. Third, the pricing and availability regions of the Tinker platform are not yet clear, which will affect developers' willingness to try it.

The emergence of Inkling 2 Small shows that Thinking Machines Lab is productizing its research capabilities, and if the "self-correction" direction can be delivered in real scenarios, it would bring substantive change to workflows in coding and visual tasks.

One-line Summary

Thinking Machines Lab releases the 594B-parameter MoE model Inkling 2 Small, focused on self-correction and a visual feedback loop.

Paragraph Summary

Thinking Machines Lab released Inkling 2 Small on Product Hunt, a MoE model with 594B total parameters and 17B activated during inference, targeting long-horizon coding, visual tasks, and agentic scenarios. Its core selling point is "being able to see and refine what it builds," i.e., a generate-observe-revise closed loop, differentiating it from traditional one-shot generation models. The model offers inference and fine-tuning services through the Tinker platform, with Mira Murati and Tianyi Zhang as makers. It topped Product Hunt with 100 votes on launch day, but with only 2 comments, discussion remained limited. Its differentiation lies in the visual feedback loop and MoE cost advantages, but technical details, pricing, and direct comparisons with competitors have not been disclosed, and real-world reputation remains to be verified.