Inkling-Small: A 12B-Active MoE That Beats Its Teacher

Introduction

On July 30, 2026, Thinking Machines Lab released Inkling-Small, an open-weights mixture-of-experts model positioned as a smaller, more efficient sibling to Inkling, the company's first open model, which had shipped two weeks earlier on July 15. The headline result is not that Inkling-Small is cheaper to run than its larger counterpart, that would be expected of any smaller model, but that it outperforms Inkling on reasoning benchmarks despite having a fraction of the active parameters. That outcome came from a specific training pipeline: distilling from an on-policy teacher, followed by targeted reinforcement learning, rather than simply scaling down the original architecture.

Feature Overview

Inkling-Small is a mixture-of-experts transformer with 276 billion total parameters and 12 billion active per token, alongside a 1-million-token context window. It was trained on Nvidia's GB300 NVL72 systems. Thinking Machines describes the model's construction in two stages. First, the company applied a refined pre-training data composition and ML recipe, improvements developed after training on the larger Inkling model had already begun, and used them to post-train an earlier Inkling-Small checkpoint. Second, that checkpoint went through on-policy distillation, using the larger Inkling model as a teacher, followed by roughly two weeks of reinforcement learning focused on agentic coding tasks.

Architecturally, Inkling-Small is natively multimodal and encoder-free, meaning it reasons directly over audio and image inputs rather than routing them through a separate vision or audio encoder module bolted onto a text backbone. Outputs remain text-based, including code and structured data. The model also supports variable thinking effort, letting a caller trade off inference compute against response quality on a given request rather than being locked into one fixed reasoning budget.

On Thinking Machines' own reported benchmarks, Inkling-Small scores 31.6% on Humanity's Last Exam versus 29.7% for the larger Inkling, and exceeds 80% on SWE-Bench Verified. The company also describes it as competitive on token efficiency against open-weight models in its size class. These are single-vendor figures; no independent lab has published a replication of them as of this release.

For context, the parent model, Inkling, shipped on July 15, 2026, with 975 billion total parameters and roughly 41 billion active, trained on 45 trillion tokens spanning text, image, audio, and video, with the same 1-million-token context window. Inkling was released under an Apache 2.0 license with open weights. Thinking Machines was explicit at Inkling's launch that it is "not the strongest overall model available today, open or closed," positioning it instead around well-rounded, balanced performance rather than leaderboard-topping scores in any one category.

Usability Analysis

Inkling-Small's weights are published on Hugging Face, and Thinking Machines offers fine-tuning through its Tinker platform, with a Tinker Playground interface for chatting with the model across text, image, and audio inputs. Output token pricing through Tinker is listed at $1.20 per million tokens. That combination, open weights plus a managed fine-tuning and hosting layer, is the core of Thinking Machines' business model: the company does not monetize the model weights themselves through a metered API, and generates revenue instead through Tinker's training and hosting services.

That structure targets a fairly specific user: teams that want to fine-tune a capable base model for a domain-specific task rather than call a general-purpose API. One demonstrated case comes from a collaboration with Bridgewater Associates, where a fine-tuned version of an open Thinking Machines model reportedly reached 84.7% on financial reasoning tests at about one-fourteenth the running cost of the comparison system. That result illustrates the intended path, customize a smaller open model for a narrow task, rather than defaulting to the largest available general-purpose model. It is a strategy that assumes the customer already has the ML expertise to do that fine-tuning work, or the budget to have Thinking Machines do it for them.

Pros and Cons

Pros

1. Inkling-Small outperforms its own larger teacher model on reasoning benchmarks (31.6% vs. 29.7% on Humanity's Last Exam) despite having roughly a third of the active parameters, a notable result for an on-policy distillation approach.
2. Encoder-free native multimodality lets the model reason directly over audio and image inputs without a bolted-on encoder module.
3. Variable thinking effort gives deployers a way to tune compute cost against response quality per request.
4. Open weights on Hugging Face, combined with Tinker for fine-tuning and hosting, gives technical teams a clear path to customize the model rather than being limited to a fixed API.
5. SWE-Bench Verified performance above 80% suggests strong practical coding capability, reinforced by roughly two weeks of agentic-coding-focused reinforcement learning in its training pipeline.

Cons

1. All benchmark figures (Humanity's Last Exam, SWE-Bench Verified, token efficiency claims) are self-reported by Thinking Machines; none have been independently replicated as of this release.
2. Inkling-Small's own license has not been separately confirmed as Apache 2.0; only the parent Inkling model's Apache 2.0 terms are officially documented, so Inkling-Small should be described as open weights rather than assumed to carry identical licensing.
3. The customize-rather-than-scale approach depends on users having real machine learning expertise, or budget for Thinking Machines' own fine-tuning services, to get value beyond the base model.
4. At 276 billion total parameters, Inkling-Small still requires meaningful infrastructure to self-host, even though it is far lighter than the 975-billion-parameter Inkling.

Outlook

Inkling-Small is a direct test of Thinking Machines' founding thesis: that a smaller, fine-tunable model, built through targeted distillation and reinforcement learning rather than raw scale, can match or beat a larger sibling on the tasks that matter to a given customer. The fact that a 12-billion-active-parameter model outperformed its 41-billion-active-parameter teacher on reasoning benchmarks is a meaningful data point for that thesis, though it rests on the company's own numbers for now.

The Bridgewater case, an 84.7% score on financial reasoning at roughly one-fourteenth the cost of the comparison system, is the kind of evidence that would need to recur across other domains and other customers before the customize-don't-scale approach can be judged as broadly viable rather than a single favorable case study. With around 200 employees and a path to market of roughly nine months from founding, Thinking Machines' bet is that revenue from Tinker's fine-tuning and hosting services, not metered access to the base weights, is where the business sustains itself. Whether that model scales past early flagship partnerships is the open question.

Conclusion

Inkling-Small is a genuinely interesting result: a smaller, open-weights model trained through on-policy distillation from its own larger teacher, that then outperforms that teacher on reasoning benchmarks. That outcome, plus native encoder-free multimodality and variable thinking effort, makes it a reasonable candidate for teams already invested in fine-tuning open models for specific tasks. The benchmark claims are vendor-reported and not yet independently verified, and the approach assumes real ML capacity on the customer's side, so it is best suited to technical teams and enterprises with a concrete fine-tuning use case, rather than a general drop-in replacement for a hosted frontier API.

Editor's Verdict

Inkling-Small: A 12B-Active MoE That Beats Its Teacher earns a solid recommendation within the other llm space.

The strongest case for paying attention is outperforms its own larger teacher model on reasoning benchmarks despite roughly a third of the active parameters, which raises the bar for what readers should now expect from peers in this space. Reinforcing that, encoder-free native multimodality reasons directly over audio and image inputs adds practical value rather than just headline appeal. The broader signal worth registering is straightforward: inkling-Small (12B active) outperforms its own larger teacher, Inkling (41B active), on Humanity's Last Exam — 31.6% vs. 29.7% — a notable result for an on-policy distillation pipeline. On the other side of the ledger, all benchmark figures are self-reported by Thinking Machines and not yet independently replicated is a real constraint, not a marketing footnote, and it should factor into any serious decision. Layered on top of that, inkling-Small's specific license has not been separately confirmed as Apache 2.0; it should be described as open weights narrows the set of teams for whom this is an obvious yes.

For multi-model deployment teams, cost-conscious operators, and developers willing to evaluate beyond the major labs, this is a serious evaluation candidate, not just a curiosity to bookmark. For everyone else, the safer posture is to monitor coverage and revisit once the use cases that matter to your team are demonstrated in the wild.