Inkling-Small評測：Mira Murati開源模型以小勝大 | Inkling-Small Review: Small Open Model Beats Big Sibling

8月 03, 2026

By Kit 小克 | AI Tool Observer | 2026-08-04

🇹🇼 Inkling-Small評測：Mira Murati開源模型以小勝大

Inkling-Small 是前 OpenAI CTO Mira Murati 創辦的新創 Thinking Machines Lab，在2026年8月初推出的開源混合專家（MoE）模型，總參數276B、實際啟用僅12B，卻幾乎打平自家4倍體積的旗艦模型 Inkling，是本週最受關注的開源AI話題之一。

什麼是 Inkling-Small？

Inkling-Small 是 Thinking Machines Lab 在推出975B參數的 Inkling 僅兩週後再度發布的縮小版模型，採用 Apache 2.0 授權完全開源，任何人都能免費下載、微調、商用，不受供應商綁定限制。

效能表現如何？

根據第三方評測機構 Artificial Analysis 的智能指數，Inkling-Small 與 Inkling 僅差一分。在 GPQA Diamond 測驗中，Inkling-Small 拿下88.3%，甚至略勝 Inkling 的87.2%；在 HLE（人類最後一考）工具輔助測驗中也以46.6%小贏46.0%。

推理與程式能力：Inkling-Small 在多項 agentic coding 基準上反而超越大自己四倍的 Inkling
知識覆蓋面：Inkling 仍保有優勢，畢竟參數規模較大、訓練涵蓋更廣
多模態：兩者皆在45兆 token 上訓練，涵蓋文字、圖像、音訊、影片，並支援百萬 token 上下文

怎麼使用 Inkling-Small？

Inkling-Small 的權重已在 Hugging Face 開放下載，也能透過 Thinking Machines 的微調平台 Tinker 直接用 LoRA 客製化。公司不靠 API 計量收費，獲利模式來自 Tinker 訂閱，這與 OpenAI、Anthropic 的閉源訂閱模式明顯不同。

Inkling-Small 值不值得用？

如果你在意開源授權與可自行部署，Inkling-Small 是目前美系開源模型中少見能打進第一梯隊智能指數的選項，尤其「小模型打平大模型」這件事本身就證明了 Thinking Machines 在效率化訓練上的技術力。但這只是新公司第二個公開模型，社群生態與工具鏈成熟度都還在起步，跟 Llama、Qwen 這種已經跑了好幾代的開源家族相比，還需要時間累積實戰案例。

好不好用，試了才知道。

🇺🇸 Inkling-Small Review: Small Open Model Beats Big Sibling

Inkling-Small is the new open-weights model from Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati. Released in early August 2026, it packs 276B total parameters (only 12B active) yet nearly matches the performance of its own 975B-parameter flagship, Inkling — making it one of the most talked-about open-source AI releases this week.

What Is Inkling-Small?

Inkling-Small is a mixture-of-experts (MoE) model Thinking Machines shipped just two weeks after its first model, Inkling. It is released under the Apache 2.0 license, meaning anyone can download, fine-tune, and deploy it commercially with no vendor lock-in.

How Does Inkling-Small Perform?

On the third-party Artificial Analysis Intelligence Index, Inkling-Small trails Inkling by less than a single point despite being a quarter of the size. It actually edges out Inkling on GPQA Diamond (88.3% vs 87.2%) and on HLE with tools (46.6% vs 46.0%).

Reasoning and coding: Inkling-Small beats its four-times-larger sibling on several agentic coding benchmarks
Knowledge coverage: Inkling still holds an edge here, thanks to its larger parameter count and broader training
Multimodality: Both models were trained on 45 trillion tokens spanning text, image, audio, and video, and support up to a 1M-token context window

How Do You Use Inkling-Small?

Weights are already live on Hugging Face, and the model is fine-tunable through Thinking Machines' own platform, Tinker, using LoRA. Notably, the company does not monetize Inkling-Small through metered API access — revenue comes from Tinker subscriptions instead, a different model from the closed, subscription-based approach at OpenAI and Anthropic.

Is Inkling-Small Worth Using?

If you care about open licensing and self-hosting, Inkling-Small is one of the few US-built open models currently competitive at the top of the intelligence index — the fact that a quarter-sized model nearly matches its flagship says something real about Thinking Machines training efficiency. That said, this is only the startup's second public model. Compared to battle-tested open families like Llama or Qwen, the tooling and community ecosystem around it are still young.

好不好用，試了才知道 — you will not know if it fits your stack until you try it.

Sources / 資料來源

常見問題 FAQ

Inkling-Small 跟 Inkling 差在哪？

Inkling-Small 是276B總參數（12B啟用）的縮小版模型，效能與975B的Inkling幾乎相同，在部分推理與程式測驗上甚至更勝一籌。

Inkling-Small 可以免費商用嗎？

可以，Inkling-Small 採用 Apache 2.0 授權，允許免費下載、修改與商業部署，不受供應商綁定。

要去哪裡下載 Inkling-Small？

權重已上架 Hugging Face，也可透過 Thinking Machines 的 Tinker 平台直接進行 LoRA 微調。

Inkling-Small 是哪家公司推出的？

由前 OpenAI CTO Mira Murati 創辦的新創 Thinking Machines Lab 於2026年8月推出。

延伸閱讀 / Related Articles

AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

搜尋此網誌

AI小貼士