Thinking Machines Lab, the San Francisco startup founded by former OpenAI chief technology officer Mira Murati, has released Inkling-Small, a 276-billion-parameter open-weights model that arrives barely two weeks after the company shipped its first model, the far larger Inkling. The headline is not the size but the ratio: a model with less than a third of its predecessor's parameters lands within a single point of it on a widely watched capability ranking, and pulls ahead on several coding and agentic tests. For a company that has staked its identity on the argument that bigger is not automatically better, it is a pointed demonstration.
Inkling-Small is a mixture-of-experts system with 276 billion total parameters but only about 12 billion active for any given token, released under the permissive Apache 2.0 license. It accepts text, image, and audio inputs, returns text, and supports a context window of up to one million tokens. On the Artificial Analysis Intelligence Index, it scored 40 against 41 for the flagship Inkling, which carries 975 billion total parameters and 41 billion active. Artificial Analysis noted that no open-weight model of the same size or smaller has scored higher.
.jpg){kind=link}
The benchmark picture is where the release gets interesting. On SWE-bench Verified, a test of real software-engineering fixes, Inkling-Small posted 80.2 percent to the larger model's 77.6 percent. On Terminal Bench 2.1 it reached 64.7 percent versus 63.8 percent. And on AA-Briefcase, Artificial Analysis's long-horizon agentic knowledge-work benchmark, the smaller model recorded 917 Elo against 839 for its sibling. In other words, the cheaper model is not merely close; on the tasks enterprises increasingly care about, it is sometimes better.
Why It Matters
The economics of running a model, not training it, are what most businesses actually pay for. A 276-billion-parameter mixture-of-experts model that activates only 12 billion parameters per token can be served on dramatically less hardware than a near-trillion-parameter system, which translates directly into lower cost per request, faster responses, and the option to run the model on infrastructure a mid-sized company can realistically afford. When that smaller model matches or beats the large one on coding and agentic work, the calculus for buyers shifts.
There is also the license to consider. Apache 2.0 open weights mean companies can download Inkling-Small, fine-tune it on proprietary data, and deploy it inside their own environments without sending anything to an external provider. For regulated industries and firms wary of vendor lock-in, that combination of open licensing and modest compute requirements is arguably more consequential than a point or two on a leaderboard. It reframes the competitive question from which lab has the smartest model to which lab gives customers the most control over a capable one.
Thinking Machines has been explicit that this is the bet. Rather than chase a single frontier system meant to do everything, the company argues that most real deployments are better served by efficient, adaptable models that teams can shape to their own needs. Inkling-Small is the clearest evidence yet of what that philosophy looks like when it ships.
The Reaction
Among developers, the release landed less as a spectacle and more as a practical gift. An open-weights model that rivals a much larger sibling on engineering benchmarks is exactly the kind of tool that independent builders, academic labs, and startups can put to work immediately, without negotiating enterprise contracts or budgeting for frontier-scale inference bills. Within the open-source community, the appeal of a model that can be inspected, modified, and self-hosted is a recurring theme, and Inkling-Small fits it neatly.
The comparisons were also pointed. Analysts noted that the strongest American Apache 2.0 open model now comes from a company barely a year into shipping products, rather than from the established giants whose flagship systems remain closed. That framing matters for morale in the open ecosystem, which has spent much of the past two years watching the most capable models disappear behind paywalls and usage tiers. A credible, openly licensed alternative from a high-profile lab is a signal that the open track is still competitive.
What Comes Next
The immediate question is cadence. Two model releases in roughly two weeks suggests Thinking Machines intends to iterate quickly, and the pattern of a large flagship followed by a distilled, more efficient sibling hints at a product line rather than a one-off. If the company can keep compressing capability into smaller footprints, it will be pushing directly on the cost curve that determines how widely AI actually gets deployed.
From its base in San Francisco, Thinking Machines is competing in the same talent-dense corridor as OpenAI, Anthropic, and Google, all of which are refining their own strategies around open versus closed releases and around the balance between raw capability and serving cost. Expect the larger labs to respond, whether by releasing more open weights of their own or by sharpening the price and efficiency of their hosted offerings. The watchers to keep an eye on are the benchmark trackers and the enterprise buyers, because their verdicts, more than any launch-day claim, will decide whether the efficient-and-open thesis holds.
Closing Thoughts
Inkling-Small is a small model making a large argument. Its numbers suggest that the industry's reflex to equate parameter count with intelligence is due for revision, and that the more durable advantage may lie in how efficiently and openly a capable model can be delivered. If that view is right, the next phase of the AI race will be won less on the size of the model and more on the terms under which people can use it.
None of this settles the debate. Frontier-scale systems still lead on the hardest reasoning tasks, and a single benchmark cycle is a thin basis for sweeping conclusions. But Inkling-Small is a concrete, downloadable data point in favor of a leaner future, and it comes from a lab that people are watching closely. For businesses weighing how to adopt AI without surrendering cost control or data governance, that is a development worth tracking.
한글 요약
미라 무라티 전 오픈AI 최고기술책임자가 설립한 샌프란시스코 스타트업 씽킹 머신스 랩이 오픈 웨이트 모델 '잉클링-스몰(Inkling-Small)'을 공개했습니다. 총 2,760억 파라미터에 토큰당 약 120억 파라미터만 활성화하는 전문가 혼합(MoE) 구조이며, 아파치 2.0 라이선스로 배포됩니다. 텍스트·이미지·오디오 입력을 받고 최대 100만 토큰의 컨텍스트를 지원합니다.
핵심은 크기가 아니라 효율입니다. 파라미터가 3분의 1도 안 되는데도, 9,750억 파라미터의 상위 모델 '잉클링'과 아티피셜 애널리시스 인텔리전스 인덱스에서 40 대 41로 단 1점 차이를 냈고, 실제 소프트웨어 수정 능력을 재는 SWE-bench Verified(80.2% 대 77.6%)와 장기 에이전트 작업 벤치마크에서는 오히려 상위 모델을 앞섰습니다. 활성 파라미터가 적을수록 추론 비용과 응답 속도, 필요한 하드웨어 규모가 줄어들기 때문에 기업 입장에서는 도입 문턱이 크게 낮아집니다.
오픈 라이선스와 낮은 구동 비용의 결합은 규제 산업이나 데이터 통제가 중요한 기업에 특히 의미가 큽니다. 씽킹 머신스는 '모든 것을 하는 하나의 거대 모델'보다 팀이 직접 다듬어 쓸 수 있는 효율적 모델이 대부분의 실제 배포에 더 적합하다는 전략을 내세워 왔고, 잉클링-스몰은 그 논리를 가장 분명하게 보여 준 사례입니다. 다만 최고난도 추론에서는 여전히 초대형 모델이 앞서며, 한 번의 벤치마크로 결론을 내리기엔 이르다는 점도 함께 짚어 둘 필요가 있습니다.