Inkling-Small評測：Mira Murati開源模型以小搏大 | Inkling-Small Review: Open-Weight AI Punches Above Weight

By Kit 小克 | AI Tool Observer | 2026-08-10

🇹🇼 Inkling-Small評測：Mira Murati開源模型以小搏大

Inkling-Small是什麼？Mira Murati新創的「以小搏大」實驗

Inkling-Small是前OpenAI技術長Mira Murati創辦的Thinking Machines Lab於2026年7月30日推出的開源多模態AI模型，距離母模型Inkling發表僅隔15天。這次評測的重點：Inkling-Small用僅2760億參數（啟用參數僅120億），做到接近甚至部分超越975億參數母模型的效能，體積只有母模型的四分之一左右，卻沒有明顯犧牲推理與程式碼能力。

技術規格與效能實測

參數規模：2760億總參數，僅啟用120億（Mixture-of-Experts架構）
多模態輸入：支援文字、圖像、音訊輸入，輸出為文字
授權：Apache 2.0，可免費商用、可微調
基準測試：在HLE（Humanity's Last Exam）等指標上，Inkling-Small在部分項目甚至贏過體積大它三倍多的母模型Inkling

這種「小模型打敗大模型」的現象在2026年並不新鮮——蒸餾與資料配方優化已經是主流打法，但Inkling-Small的意義在於它是開源權重釋出，任何團隊都能下載到Hugging Face上直接部署或微調，不用被鎖在單一供應商的API裡。

定價與怎麼用

透過OpenRouter，Inkling-Small的API定價約為每百萬輸入token 0.45美元、輸出1.20美元，上市初期還有五折優惠。相比動輒每百萬token兩三美元起跳的閉源前沿模型，這個價位對中小型團隊或高頻呼叫的應用相當友善。企業也可以透過Thinking Machines自家的Tinker微調工具，針對特定任務（例如金融分析，避險基金Bridgewater就是早期客戶）客製化模型行為。

值得注意的商業邏輯

Thinking Machines在2025年就以「還沒有產品」的狀態募到12億美元估值的20億美元種子輪，外界一度質疑這筆錢是不是純粹押注團隊聲望。Inkling與Inkling-Small的推出算是第一次拿出真正能用的東西。公司押注的邏輯很清楚：企業要的不是「租」一個黑盒子模型，而是能夠擁有並客製化的AI——這跟OpenAI、Anthropic走的閉源訂閱路線是完全不同的賭注。

小克的實測建議

如果你在做需要自架、微調或成本敏感的應用（例如客服機器人、內部知識庫問答），Inkling-Small值得下載測試——它的開源與價格優勢是實打實的。但如果你需要的是最頂尖的推理能力，目前主流閉源模型（GPT-5.5、Fable 5等）仍然領先。Inkling-Small更適合定位在「夠好又便宜還能自己掌控」的中間地帶。

好不好用，試了才知道。

🇺🇸 Inkling-Small Review: Open-Weight AI Punches Above Weight

Inkling-Small Review: Can a Quarter-Sized Model Really Match Its Parent?

Inkling-Small is the open-weight multimodal model that Mira Murati'''s Thinking Machines Lab released on July 30, 2026 — just 15 days after its parent model Inkling. The headline claim: at roughly a quarter of the parameter count, Inkling-Small matches or even beats Inkling on several benchmarks, without a meaningful hit to reasoning or coding quality.

Specs and What the Benchmarks Actually Show

Parameters: 276B total, only 12B active per token (Mixture-of-Experts)
Multimodal input: text, image, and audio in; text out
License: Apache 2.0 — free for commercial use and fine-tuning
Benchmarks: on tests like Humanity'''s Last Exam (HLE), Inkling-Small edges out the 975B-parameter Inkling on some tasks despite being roughly a third the active-parameter size

Beating a bigger sibling model isn'''t new in 2026 — better data recipes and distillation are now standard practice. What makes Inkling-Small notable is that it ships as open weights on Hugging Face, so any team can self-host or fine-tune it without being locked into one vendor'''s API.

Pricing and How to Access It

Through OpenRouter, Inkling-Small runs about $0.45 per million input tokens and $1.20 per million output tokens, with a launch discount cutting that further. Compared to frontier closed models that often start at $2-3 per million tokens, this pricing is friendly for smaller teams or high-volume workloads. Enterprises can also fine-tune via Thinking Machines''' own Tinker API — hedge fund Bridgewater Associates is an early customer using it for financial-analysis tasks.

The Business Bet Behind It

Thinking Machines raised a $2 billion seed round at a $12 billion valuation in 2025 before shipping a single product, drawing skepticism that the money was chasing team pedigree rather than substance. Inkling and Inkling-Small are the company'''s first real answer to that criticism. The bet is straightforward: enterprises don'''t want to rent a black-box model, they want to own and customize one — a different wager than the closed, subscription-based paths OpenAI and Anthropic are running.

Kit'''s Take

If you'''re building something that needs self-hosting, fine-tuning, or tight cost control — internal knowledge bases, customer support bots — Inkling-Small is worth downloading and testing. The open license and pricing are real advantages. If you need absolute top-tier reasoning, closed frontier models like GPT-5.5 or Fable 5 still lead. Inkling-Small'''s sweet spot is "good enough, cheap, and yours to control."

好不好用，試了才知道。

Sources / 資料來源

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

Claude Code 實測：AI 幫你寫程式到底行不行？ | Claude Code Review: Can AI Really Code for You?

3月 26, 2026

閱讀完整內容

搜尋此網誌

AI小貼士